MODERN OPTIMIZATION OF ROTATING ELECTRIC MACHINES · Place of Birth: V¨ocklabruck, Austria ......
Transcript of MODERN OPTIMIZATION OF ROTATING ELECTRIC MACHINES · Place of Birth: V¨ocklabruck, Austria ......
MODERN OPTIMIZATIONOF
ROTATING ELECTRIC MACHINES
Gerd BramerdorferDepartment of Electrical Drives and Power Electronics
Johannes Kepler University Linz Linz Austriagerdbramerdorferjkuat
Antti LehikoinenDepartment of Electrical Engineering
Aaalto Universityanttilehikoinenaaltofi
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Ass-Prof Dipl-Ing Dr Gerd Bramerdorfer
Date of Birth February 7 1981
Place of Birth Vocklabruck Austria
Nationality Austrian
Email gerdbramerdorferjkuat
Phone +43(0)7322468-6431
GoogleScholar GerdGoogleScholar
LinkedIn GerdLinkedIn
ResearchGate GerdResearchGate
Gerd Bramerdorfer September 2 2018 Slide 2
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Short Biography
Gerd Bramerdorfer is an Assistant Professor with Johannes KeplerUniversity Linz and a senior researcher and project leader at the LinzCenter of Mechatronics
He received the Dipl-Ing degree in Mechatronics in 2007 and thePhD degree in Electrical Engineering in 2014 both from JohannesKepler University Linz Linz Austria
Since 2007 he has been with the Department of Electrical Drivesand Power Electronics Johannes Kepler University Linz where hewas involved in various research projects and held classes in the fieldof electrical machines and actuators and power electronics
Gerd Bramerdorfer September 2 2018 Slide 3
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Short Biography contrsquod
Gerd Bramerdorfer is a member of IEEE IEEE Industrial ElectronicsSociety and its Electric Machines Technical Committee and theIEEE Industry Applications Society
He constantly serves for the scientific community eg as guesteditor for a special section in IEEE Transactions on IndustrialElectronics called rdquoOptimization of Electric Machine Designsrdquo astrack co-chair topic chair by organizing special sessions and as areviewer for journals and conferences He is an Associate Editor ofthe IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
Publications etc can be found onlineGerd Bramerdorfer September 2 2018 Slide 4
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Research stays
012016ndash032016 Guest researcher at the University of Wis-consin Madison US and the WisconsinElectric Machines and Power ElectronicsConsortium (WEMPEC)Invited by Prof Bulent Sarlioglu
032016ndash052016 Guest researcher at the Politecnico diTorino Dipartimento di Energia ItalyInvited by Prof Andrea Cavagnino
Gerd Bramerdorfer September 2 2018 Slide 5
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Main Research Interests
Design and modeling of electric machines and drives
Multi-objective optimization of mechatronic systems
Sensitivity and tolerance analyses
Robust optimization optimization for high reliability
Material characterization for electric machine modeling(1D- and 2D-characterization impact of manufacturing)
Nonlinear optimal control of electric machines
Magnetic levitation and magnetic bearing technology
Gerd Bramerdorfer September 2 2018 Slide 6
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation Outline
Gerd Bramerdorfer September 2 2018 Slide 7
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
Gerd Bramerdorfer September 2 2018 Slide 8
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Ass-Prof Dipl-Ing Dr Gerd Bramerdorfer
Date of Birth February 7 1981
Place of Birth Vocklabruck Austria
Nationality Austrian
Email gerdbramerdorferjkuat
Phone +43(0)7322468-6431
GoogleScholar GerdGoogleScholar
LinkedIn GerdLinkedIn
ResearchGate GerdResearchGate
Gerd Bramerdorfer September 2 2018 Slide 2
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Short Biography
Gerd Bramerdorfer is an Assistant Professor with Johannes KeplerUniversity Linz and a senior researcher and project leader at the LinzCenter of Mechatronics
He received the Dipl-Ing degree in Mechatronics in 2007 and thePhD degree in Electrical Engineering in 2014 both from JohannesKepler University Linz Linz Austria
Since 2007 he has been with the Department of Electrical Drivesand Power Electronics Johannes Kepler University Linz where hewas involved in various research projects and held classes in the fieldof electrical machines and actuators and power electronics
Gerd Bramerdorfer September 2 2018 Slide 3
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Short Biography contrsquod
Gerd Bramerdorfer is a member of IEEE IEEE Industrial ElectronicsSociety and its Electric Machines Technical Committee and theIEEE Industry Applications Society
He constantly serves for the scientific community eg as guesteditor for a special section in IEEE Transactions on IndustrialElectronics called rdquoOptimization of Electric Machine Designsrdquo astrack co-chair topic chair by organizing special sessions and as areviewer for journals and conferences He is an Associate Editor ofthe IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
Publications etc can be found onlineGerd Bramerdorfer September 2 2018 Slide 4
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Research stays
012016ndash032016 Guest researcher at the University of Wis-consin Madison US and the WisconsinElectric Machines and Power ElectronicsConsortium (WEMPEC)Invited by Prof Bulent Sarlioglu
032016ndash052016 Guest researcher at the Politecnico diTorino Dipartimento di Energia ItalyInvited by Prof Andrea Cavagnino
Gerd Bramerdorfer September 2 2018 Slide 5
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Main Research Interests
Design and modeling of electric machines and drives
Multi-objective optimization of mechatronic systems
Sensitivity and tolerance analyses
Robust optimization optimization for high reliability
Material characterization for electric machine modeling(1D- and 2D-characterization impact of manufacturing)
Nonlinear optimal control of electric machines
Magnetic levitation and magnetic bearing technology
Gerd Bramerdorfer September 2 2018 Slide 6
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation Outline
Gerd Bramerdorfer September 2 2018 Slide 7
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
Gerd Bramerdorfer September 2 2018 Slide 8
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Short Biography
Gerd Bramerdorfer is an Assistant Professor with Johannes KeplerUniversity Linz and a senior researcher and project leader at the LinzCenter of Mechatronics
He received the Dipl-Ing degree in Mechatronics in 2007 and thePhD degree in Electrical Engineering in 2014 both from JohannesKepler University Linz Linz Austria
Since 2007 he has been with the Department of Electrical Drivesand Power Electronics Johannes Kepler University Linz where hewas involved in various research projects and held classes in the fieldof electrical machines and actuators and power electronics
Gerd Bramerdorfer September 2 2018 Slide 3
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Short Biography contrsquod
Gerd Bramerdorfer is a member of IEEE IEEE Industrial ElectronicsSociety and its Electric Machines Technical Committee and theIEEE Industry Applications Society
He constantly serves for the scientific community eg as guesteditor for a special section in IEEE Transactions on IndustrialElectronics called rdquoOptimization of Electric Machine Designsrdquo astrack co-chair topic chair by organizing special sessions and as areviewer for journals and conferences He is an Associate Editor ofthe IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
Publications etc can be found onlineGerd Bramerdorfer September 2 2018 Slide 4
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Research stays
012016ndash032016 Guest researcher at the University of Wis-consin Madison US and the WisconsinElectric Machines and Power ElectronicsConsortium (WEMPEC)Invited by Prof Bulent Sarlioglu
032016ndash052016 Guest researcher at the Politecnico diTorino Dipartimento di Energia ItalyInvited by Prof Andrea Cavagnino
Gerd Bramerdorfer September 2 2018 Slide 5
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Main Research Interests
Design and modeling of electric machines and drives
Multi-objective optimization of mechatronic systems
Sensitivity and tolerance analyses
Robust optimization optimization for high reliability
Material characterization for electric machine modeling(1D- and 2D-characterization impact of manufacturing)
Nonlinear optimal control of electric machines
Magnetic levitation and magnetic bearing technology
Gerd Bramerdorfer September 2 2018 Slide 6
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation Outline
Gerd Bramerdorfer September 2 2018 Slide 7
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
Gerd Bramerdorfer September 2 2018 Slide 8
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Short Biography contrsquod
Gerd Bramerdorfer is a member of IEEE IEEE Industrial ElectronicsSociety and its Electric Machines Technical Committee and theIEEE Industry Applications Society
He constantly serves for the scientific community eg as guesteditor for a special section in IEEE Transactions on IndustrialElectronics called rdquoOptimization of Electric Machine Designsrdquo astrack co-chair topic chair by organizing special sessions and as areviewer for journals and conferences He is an Associate Editor ofthe IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
Publications etc can be found onlineGerd Bramerdorfer September 2 2018 Slide 4
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Research stays
012016ndash032016 Guest researcher at the University of Wis-consin Madison US and the WisconsinElectric Machines and Power ElectronicsConsortium (WEMPEC)Invited by Prof Bulent Sarlioglu
032016ndash052016 Guest researcher at the Politecnico diTorino Dipartimento di Energia ItalyInvited by Prof Andrea Cavagnino
Gerd Bramerdorfer September 2 2018 Slide 5
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Main Research Interests
Design and modeling of electric machines and drives
Multi-objective optimization of mechatronic systems
Sensitivity and tolerance analyses
Robust optimization optimization for high reliability
Material characterization for electric machine modeling(1D- and 2D-characterization impact of manufacturing)
Nonlinear optimal control of electric machines
Magnetic levitation and magnetic bearing technology
Gerd Bramerdorfer September 2 2018 Slide 6
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation Outline
Gerd Bramerdorfer September 2 2018 Slide 7
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
Gerd Bramerdorfer September 2 2018 Slide 8
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Research stays
012016ndash032016 Guest researcher at the University of Wis-consin Madison US and the WisconsinElectric Machines and Power ElectronicsConsortium (WEMPEC)Invited by Prof Bulent Sarlioglu
032016ndash052016 Guest researcher at the Politecnico diTorino Dipartimento di Energia ItalyInvited by Prof Andrea Cavagnino
Gerd Bramerdorfer September 2 2018 Slide 5
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Main Research Interests
Design and modeling of electric machines and drives
Multi-objective optimization of mechatronic systems
Sensitivity and tolerance analyses
Robust optimization optimization for high reliability
Material characterization for electric machine modeling(1D- and 2D-characterization impact of manufacturing)
Nonlinear optimal control of electric machines
Magnetic levitation and magnetic bearing technology
Gerd Bramerdorfer September 2 2018 Slide 6
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation Outline
Gerd Bramerdorfer September 2 2018 Slide 7
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
Gerd Bramerdorfer September 2 2018 Slide 8
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Main Research Interests
Design and modeling of electric machines and drives
Multi-objective optimization of mechatronic systems
Sensitivity and tolerance analyses
Robust optimization optimization for high reliability
Material characterization for electric machine modeling(1D- and 2D-characterization impact of manufacturing)
Nonlinear optimal control of electric machines
Magnetic levitation and magnetic bearing technology
Gerd Bramerdorfer September 2 2018 Slide 6
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation Outline
Gerd Bramerdorfer September 2 2018 Slide 7
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
Gerd Bramerdorfer September 2 2018 Slide 8
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation Outline
Gerd Bramerdorfer September 2 2018 Slide 7
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
Gerd Bramerdorfer September 2 2018 Slide 8
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
Gerd Bramerdorfer September 2 2018 Slide 8
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 9
Many design parameters
diametersmagnet height
Multiple objectives
efficiencycosttorque rippleripple of DC bus currentminimize capacitancetorque density
Constraints
Motor sizeSize of PCBThd
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Why should we deal with optimization techniques
Because we want to have the best machine design the bestpower electronics the best control or the best overall solutionfor a particular application
Why cannot we trust our feelings and knowledge in electrical
engineering
Because we usually deal with nonlinear problems that featureseveral (conflicting) objectives many design parameters andtypically some constraints
But do not be disappointed we will see later onknowledge in electrical engineering is priceless whenit comes to optimization problems in our field
Gerd Bramerdorfer September 2 2018 Slide 10
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Anyway in order to select a suitable optimization technique andto apply it in an appropriate way some knowledge aboutoptimization is very valuable
Gerd Bramerdorfer September 2 2018 Slide 11
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 12
Expert in
optimization
Problem formulation(Math Engin)objectives constraints
geometry parameter definition
Optimization techniques (Math)Deterministic (eg gradient-based) or
stochastic (eg evolutionary algorithms)approaches
Expert in
electric machine deviceoptimization
Classical machine device designmaterials electromagnetics winding
Multi-physics aspectsthermal rotor dynamics
Power electronicsand control
Machine device evaluation(Math Engin))analytical finite elements
Machine device modeling(Math Engin)analytical meta-models
Aspects of automation(Comp science)cluster environment
error handling quality measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
No Goals Restrictions
Cover all topics in very much detail and consider all differentelectromechanical devicesWe will mainly focus on rotating electric machines
Consider all optimization techniques availableWe will discuss types and characteristics of techniques andclassify them but the field is very emerging and multitudinous
Have a universally applicable approach for all problemsProblemsrsquo variety is incredibly high
Gerd Bramerdorfer September 2 2018 Slide 13
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Goals Focus
Get a broad overview of fields relevant to electric machine device optimization
Get a good understanding about how fields interact and whataffects (optimal) machine device design
Get prepared to solve real-world problems even though an exactequivalent problem was not considered before
Gerd Bramerdorfer September 2 2018 Slide 14
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introduction to Optimization
Optimization Techniques
Gerd Bramerdorfer September 2 2018 Slide 15
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 16
Single-objective optimization
Objective
min f (x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Maximizationmax f (x) = min minusf (x)
Obtaining the objective value for adesign by evaluating
an analytic equation (white box model)
a sophisticated model (typ black box)
eg artificial neural networks (ANN)
radial basis functions (RBF)
support vector machines (SVM)
a complex analysis which for instanceinvolves a finite element simulation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 17
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Unimodal function with single design parameter
ldquoConvex functionrdquo
Gradient-based techniquescan be applied
Single minimium exists
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 18
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
minus5 0 5 10 15 20 25 30
0
1
2
3
4
5
6
7
8
9
10
x
y
Multimodal function with single design parameter
Here just a single globalminimum exists
Gradient-based techniquescan be applied butresults higly depend onstarting point because
local minima exist
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of objective functions
Gerd Bramerdorfer September 2 2018 Slide 19
x
-5 -4 -3 -2 -1 0 1 2 3 4 5
y
-5
-4
-3
-2
-1
0
1
2
3
4
5
Multimodal function with two design parameters
Usually we deal even with more than two design parameters
Complexity rises rarr the need for sophisticated optimization techniques rises
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 20
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
alpham180degpz
hm
dri
dro
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Properties of design ldquoparametersrdquo
Gerd Bramerdorfer September 2 2018 Slide 21
Design parameters can
be single-valued continuous likemagnet height
diameter(s)
be single-valued discrete likeNumber of stator slots poles phases
Number of sheets of the stack
have binary states likeSingle-Layer winding(No double layer winding is applied)
Interior rotor(No Exterior rotor is used)
have multiple values likematerial characteristics (eg BH-curves stress-strain curves)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 22
But what if we would like to deal with
multiple objectives
(As it usually will be)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 23
Multi-objective optimization
Objective
min f (x)
f (x) =q1f1(x) + q2f2(x)+
+ qmfm(x)
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Introducing weighting factorsrArr Simplified problemrArr Single-objective techniques can
be applied
Objective space is constrained bythe problem formulation
Problematic for very differentobjectives(How to compare cogging torqueefficiency and cost)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Generalized optimization problem
Gerd Bramerdorfer September 2 2018 Slide 24
Multi-objective optimization
Objective
min (f1(x) f2(x) fm(x))
Design parameters
xT =[
x1 x2 xn
]
Constraints
g(x) le 0
h(x) = 0
Comments
Multiple objectives are optimizedat once
Objective space is not constrainedby problem formulation
But how is now optimalitydefined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Introducing the term Pareto Optimal
Gerd Bramerdorfer September 2 2018 Slide 25
ldquoNamed after Vilfredo Pareto Pareto optimality is a measure of
efficiency An outcome of a game is Pareto optimal if there is no other
outcome that makes every player at least as well off and at least one
player strictly better off That is a Pareto Optimal outcome cannot
be improved upon without hurting at least one playerrdquo
Source httpwwwgametheorynetdictionaryParetoOptimalhtml
What does this mean in practice
It says that a design A is Pareto optimal if there exists no otherdesign that is better in one objective and at least as good as design Afor the other objectives
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - non-conflictingobjectives
Gerd Bramerdorfer September 2 2018 Slide 26
0 01 02 03 04 05 06 07 08 09 105
06
07
08
09
1
relative losses
eff
icie
ncy
true Pareto front (single point)Pareto point nonminusdominated pointdominated points
true Pareto frontthe (not-known) overalloptimum optima
Pareto-point(s) non-dominated point(s)the optimum optima thatwas were already identified inthe optimization run
dominated point(s)points of the optimization runthat do not represent optima
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Pareto Optimality - conflicting objectives
Gerd Bramerdorfer September 2 2018 Slide 27
0 01 02 03 04 0505
06
07
08
09
1
relative costs
eff
icie
ncy
true Pareto front
Pareto points nonminusdominated points
dominated points
Efficiency and cost are usuallyconflicting objectivesrArr A tradeoff has to be found
If weighting factors are introduceda single-solution is obtained
Instead if a multi-objective problemis analyzed (like here) one can studythe trend of conflicting objectivesafter the optimization run andapply (intuitive) weighting factors
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space
Gerd Bramerdorfer September 2 2018 Slide 28
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Mapping of two-dimensional design spaceto two-dimensional objective space -constraints
Gerd Bramerdorfer September 2 2018 Slide 29
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization Techniques
Gradient-based approaches
Grid analysis
Techniques for a (more) efficient exploration of the design objective space
Design of experiments (DOE)
Genetic algorithms (GA)
Particle swarm optimization (PS)
Differential evolution (DE)
Gerd Bramerdorfer September 2 2018 Slide 30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
are very fast for a local search
can cause problems if multiple minima (maxima) are present
famous examples Newton Gauss-NewtonLevenberg-Marquardt
some techniques include mutation eg simulated annealing
are usually not applicable if many parameters are varied formultiple objectives
Gerd Bramerdorfer September 2 2018 Slide 31
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gradient-based approaches
Gerd Bramerdorfer September 2 2018 Slide 32
Source SyMSpace documentation
Sketch of equipotential lines and how a gradient-basedapproach works starting from x0
Starting point and step-width are crucial parameters tobe defined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Grid analysis
is a possible choice for a low number of design parameters
for high number of parameters too many designs would have tobe calculated(eg 10 steps per param n params 10n designs to beevaluated)
evaluation of a coarse grid can be a good starting point forother techniques (but usually better to use a technique of DOE)
Gerd Bramerdorfer September 2 2018 Slide 33
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments
is a theory about selecting ldquomost importantrdquo parametercombinations among all possible combinations of a grid
very often done in advance to the start of an evolutionaryalgorithm
famous examples Latin-Hypercube-sampling (LHS)Box-Behnken-sampling (BBS)
Gerd Bramerdorfer September 2 2018 Slide 34
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Design of experiments - LHS
Gerd Bramerdorfer September 2 2018 Slide 35
X1
X2
X1
X2
LHS - good vs bad configuration design space exploration
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms
inspiration a population like human beings but usually of fixedsize npop
typically the analysis is split into generations One generationinvolves the analysis of npop individuals
like in the nature the genetic algorithm has functions forselection (the strongest shall survive)
recombination crossover (generation of new individuals basedon experience)
mutation (to not get stuck in local minima)
Probably most prominent representativesNon-dominated Sorting Genetic Algorithm (NSGA-II)Strength Pareto Evolutionary Algorithm (SPEA2)
Gerd Bramerdorfer September 2 2018 Slide 36
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - definitions
Individual A particular design to be investigated eg amachine design
Population A total number of n individuals
Gene A particular property of an individualeg a binary number or a value in the range of a continuousdomain(useInteriorRotor (binary) stator outer diameter (continuousrepr))
Chromosome A set of parameters which defines an individual
Gerd Bramerdorfer September 2 2018 Slide 37
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - selection
Various selection techniques exist eg
Tournament selection Select the best nbest individuals of thepopulation for crossover
Roulette selection Chance of being picked is a function of thefitness of the individuals (often important in order to keepvariety in the population)
Gerd Bramerdorfer September 2 2018 Slide 38
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 39
parents
parents
offspring
offspring
+
+
+
+
a)
b)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - crossover mutation
- Binary examples
Gerd Bramerdorfer September 2 2018 Slide 40
1 1
1 1
1 1
0 0
0 0
1 1
1 1
0 0
0
0
0
1 0 0
0 1
parents offspring
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Particle Glowworm swarm optimizationinspiration movement of birds
a crowd of individuals follows to the direction of ldquobest possiblevalues for objectivesrdquo
velocity vector of an individual is based on its own best knownposition and the best position of the entire crowd(or just the best position of individuals within a certain rangearound that particular individual is used)
mutation is introduced by adding arbitrary velocity vectors byldquoovershootingrdquo the best position
Eg newPos = actPos + randomFactor1 lowast (entireBestPos minus actPos)++randomFactor2 lowast (ownBestPos minus actPos) + randomFactor3
Gerd Bramerdorfer September 2 2018 Slide 41
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution (DE)
Also population based
Individuals are called agents
An individual is replaced if a new individual has (a) bettervalue(s) for the fitness function(s)
Usually a new candidate is defined based on two parent vectors(a) a parent vector x directly chosen from the currentpopulation(b) a parent (mutant) vector v that is constructed according toa particular mutant strategy
Parameters Differential weight F crossover probability CR population size NP
Gerd Bramerdorfer September 2 2018 Slide 42
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common mutation operators
rand1
v = r1 + F (r2 minus r3)
rand2
v = r1 + F (r2 minus r3) + F (r4 minus r5)
current-to-rand1
v = r1 + U(r1 minus x) + F (r2 minus r3)
where ri 6= x i isin N and 1 le i le 5 are different randomly chosenindividuals F gt 0 is a control parameter and U isin [0 1] is a uniformlydistributed random value
Gerd Bramerdorfer September 2 2018 Slide 43
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Differential evolution - common crossover types
binomial
y[j ] =
v[j ] if Uj le CR or j = I
x[j ] if Uj gt CR and j 6= I
where CR isin [0 1] is a control parameter l le n is a randomly
chosen integer and Uj is an independent random variable uniformly
distributed in [01]
polynomialthe starting position k of the crossover process is chosen randomly
and the next consecutive L elements (counted in circular manner)
are chosen from v with a probability that decreases exponentially
with an increased distance from k
Gerd Bramerdorfer September 2 2018 Slide 44
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
ldquoNo free lunch theoremrdquo
In computational complexity and optimization the no free lunch
theorem is a result that states that for certain types of mathematical
problems the computational cost of finding a solution averaged
over all problems in the class is the same for any solution method
No solution therefore offers a rsquoshort cutrsquo In computing there are
circumstances in which the outputs of all procedures solving a
particular type of problem are statistically identical
Gerd Bramerdorfer September 2 2018 Slide 45
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Which algorithm should we select
Interpretation of the ldquoNo free lunch theoremrdquo
ldquoA general-purpose universal optimization strategy istheoretically impossible and the only way one strategy canoutperform another is if it is specialized to the specific problemunder considerationrdquo
ldquoA general-purpose almost-universal optimizer existstheoretically Each search algorithm performs well on almost allobjective functionsrdquo
ldquoAn algorithm may outperform another on a problem whenneither is specialized to the problem It may be that bothalgorithms are among the worst for the problemrdquo
Gerd Bramerdorfer September 2 2018 Slide 46
Source Wikipedia
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Remarks
Put all the knowledge you have about the optimization problem toyour problem formulationAny kind of design space reduction (without simultaneouslyexcluding feasible favorable solutions) is advantageous
Be aware that evolutionary algorithms due to their definitionscomprising random numbers obviously have stochastic behavior Ifyou run a problem ten times it is highly likely that you end up with10 different results
To conclude applying your skills (in engineering) is essential to getresults that are (close to) best possible solutions and are highlyreliable
We will observe this soonGerd Bramerdorfer September 2 2018 Slide 47
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Optimization ScenarioVariety of scenarios is incredibly high - an example is given
Selected ScenarioCost vs efficiency of PMSM SyncRM for particular application
Gerd Bramerdorfer September 2 2018 Slide 48
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 49
mandatory efficiencies for induction machines defined in IEC 60034-30
774
796
813
832
846
858
870
881
894
796
814
828
843
855
866
877
887
898
759
781
798
818
833
846
860
872
887
807
827
842
859
871
881
892
901
912
825
841
853
867
877
886
896
904
914
789
810
825
843
856
868
880
891
903
2 poles 4 poles 6 poles 2 poles 4 poles 6 poles
efficiency class IE-2 efficiency class IE-3
50 Hz
kW075
11
15
22
3
4
55
75
11
Excerpt of the IEC efficiency classes
Corresponding paper Comprehensive cost optimization study ofhigh-efficiency brushless synchronous machines [1]
Due to the increasing significance of energy saving multiple regulationshave been introduced by countries or international organizationsregarding the efficiency of electric machines eg the IEC 60034-30
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 50
Rare earth material price fluctuation throughout the last years
The use of PMSMs (Permanent magnet excited synchronousmachines) leads to a considerable increase of the efficiency but due tothe fluctuating price of permanent magnets especially ofNdFeB-magnets the cost of PMSMs is hardly predictable
Source ArnoldMagnetic Technologies
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Initial situation
Gerd Bramerdorfer September 2 2018 Slide 51
This work deals with a motor with P=3kW and a ratedsynchronous speed of n=1500rpm
Different motor topologies are analyzed regarding the trade offeffiency and material cost including PMSMs and SyncRMs(synchronous reluctance machines) This is done for twodifferent cost scenarios
Finally the cheapest designs fulfilling mandatory efficiencyrequirements can be determined
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Definitions
Gerd Bramerdorfer September 2 2018 Slide 52
Cost scenarios (prices in ekg)
Material Cost scenario 1 Cost scenario 2
NdFeB magnets grade N42SH 50 150
Ferrites 10 10
Aluminum 35 35
Copper 75 75
Laminated steel grade M400-50A 2 2
Motor characteristics calculated for typical working ambientconditions for long-term operation Tpm = 100C Tcual = 130C
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 53
Parametric stator design(dimensions as well as number of slots and
winding configuration changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number ofpole pairs changed during optimization)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Considered topologies
Gerd Bramerdorfer September 2 2018 Slide 54
Parametric rotor design with surfacemounted magnets (dimensions as well as
number of pole pairs changed duringoptimization)
Parametric rotor design with embeddedmagnets (dimensions as well as number of
pole pairs changed during optimizationrotor is analyzed with and without magnets)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswith PM excitation
Gerd Bramerdorfer September 2 2018 Slide 55
No load simulation isperformed
No load flux is determined
12 currents are dynamicallyset and calculated Theyare defined by the grid andthe rated current (is equalto 100 in the grid)
ilowast
q =Trate
m
2p ψpm
ilowast
q equiv 100minus100 minus75 minus50 minus25 0
0
25
50
75
100
125
150
dminusaxis current in percent
qminus
axis
curr
ent in
perc
ent
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Analysis of machine designswithout PM excitation
Gerd Bramerdorfer September 2 2018 Slide 56
18 currents (blue crosses inthe figure on the right) arecalculated
18 more currents arecalculated from FE-results(red dots in figure)
Rated current is defined bythe maximum currentdensity S the number ofstator phases and thewinding cross section Awdg
ilowast
q = m2
S Awdg ilowast
q equiv 100
minus1500 minus1000 minus500 0 500 1000 1500
minus1500
minus1000
minus500
0
500
1000
1500
dminusaxis current in AWdgs
qminus
axis
cu
rre
nt
in A
Wd
gs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 57
Cost vs efficiency is optimized for all rotor statorcombinations with aluminum or copper winding with ferriteor NdFeB-magnets
Cogging torque and torque ripple at load should be lowerthan 10
For any design under observation a nonlinear model isautomatically derived based on the FE-results [2]
The optimal current vector fulfilling the torque requirementis then separately computed for any design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Objectives and Facts
Gerd Bramerdorfer September 2 2018 Slide 58
An asynchronous evolutionary algorithm [3] based onSPEA2 [4 5] is applied to optimize all the design variants
Overall more than 15 millions of designs are investigatedusing FE-simulations (Femag [6]) The calculations aredone using a computer cluster that features 300 CPU cores
Optimization took several weeks and was performed usingMagOpt SyMSpace [7 8]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 59
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
PMSMs with NdFeB are best regarding the trade off cost vs efficiency
Next are PMSMs with ferrite magnets
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 1
Gerd Bramerdorfer September 2 2018 Slide 60
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts1
SurfMNdFeBAluCosts1
EmbeddedNdFeBCoCosts1
EmbeddedNdFeBAluCosts1
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 1
Aluminum and copper windings perform similar
Worst are synchronous reluctance machines
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Cost scenario 2
Gerd Bramerdorfer September 2 2018 Slide 61
80 825 85 875 90 925 950
20
40
60
80
100
120
140
160
180
200
Mate
rial costs
in E
uro
SurfMFerriteCo
SurfMFerriteAlu
SurfMNdFeBCoCosts2
SurfMNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
SyncRMFerriteCo
SyncRMFerriteAlu
SyncRMNoMagnetCo
SyncRMNoMagnetAlu
ηIEminus3
ηIEminus2
Efficiency in percent
Comparison of the cost vs efficiency trend for all different topologies - Cost scenario 2
Results for cost scenario 2 are similar
except that the trend for PMSMs with NdFeB magnets gets closer tothe other characteristics
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Scenario - Results - Optimal length to diameter ratio
Gerd Bramerdorfer September 2 2018 Slide 62
minus150 minus100 minus50 0 50 100 1500
20
40
60
80
100
120
Stator outer diameter in mm
Active
mo
tor
len
gth
(w
ith
ou
t e
nd
win
din
gs)
in m
m
IEminus2 minus Cost scenario 2
SurfMFerriteAlu
SurfMFerriteCo
SurfMNdFeBAluCosts2
SurfMNdFeBCoCosts2
EmbeddedNdFeBAluCosts2
EmbeddedNdFeBCoCosts2
SyncRMFerriteAlu
SyncRMFerriteCo
SyncRMNoMagnetAlu
SyncRMNoMagnetCo
Optimal size relations for IE2-requirements for all topologies - Cost scenario 2 -end winding not considered in axial length
PMSMs tend to a disc shape (3D-FE verification required)
SyncRMs tend to a more barrel shape(end winding effect is relatively reduced)
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Conclusions
Gerd Bramerdorfer September 2 2018 Slide 63
Results revealed that PMSMs featuring NdFeB-magnetsare also competitive on the marked for higher raw materialprices
SyncRMs tend to a more barrel shape than PMSMs
Single load point was considered for optimization a particular case study was evaluatedGeneral conclusions can hardly be made because resultsprobably do not generally hold for any
arbitrary torquespeed-requirementsmultiple load point evaluationspecial requirements on construction spaceetc
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation of
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 64
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments
The ldquooptimizerrdquo definitely finds the weakness of yourmodel - optimizations often need some iterations
Eg Minimization of torque ripple based on the followingdefinition
tripp = tpp
tmean
If for any reason the mean torque can have negative valuesthe torque ripple gets negative rarr the optimizer will mercilessenforce those individuals with negative mean torque as theygive the best objective values regarding torque ripple
Alternative formulation
tripp =∣
∣
∣
tpp
tmean
∣
∣
∣
Gerd Bramerdorfer September 2 2018 Slide 65
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Gerd Bramerdorfer September 2 2018 Slide 66
Optimization focused on no-load situationOptimization focused on no-load situation rarr small coil area
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - General comments (tbc)
Especially for optimizations including finite element simulations
Is your mesh definition appropriate for any kind of parametercombination (chromosome)
Gerd Bramerdorfer September 2 2018 Slide 67
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Automation - Cluster HTCondor
Computationally-expensive evaluations often require acomputer cluster
A sophisticated framework for job pre-processing scheduling evaluation post-processing is crucial
Gerd Bramerdorfer September 2 2018 Slide 68
Sourcehttpsresearchcswisceduhtcondor
Job manager of SyMSpace
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
The process of evolution - change ofPareto front with time
Gerd Bramerdorfer September 2 2018 Slide 69
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Cost
s
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Cost
s
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 70
But when
should we stop an optimization run
Letrsquos discuss quality indicators and measures
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
should determine the quality in terms of convergence and diversity
Gerd Bramerdorfer September 2 2018 Slide 71
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Gerd Bramerdorfer September 2 2018 Slide 72
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
should determine the quality in terms of convergence and diversity
Thus they are usually classified to
convergence quantifiers (eg epsilon measure general distance)
diversity quantifiers (eg spread metric generalized spread)
both convergence and diversity quantifiers (eg inversegenerational distance hypervolume indicator)
Gerd Bramerdorfer September 2 2018 Slide 73
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Gerd Bramerdorfer September 2 2018 Slide 74
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Other helpful measures
ldquoNew in Pareto frontrdquo Size of Pareto front
Distribution of Pareto variables(distinct optimal values)
Individuals per generation that cannot be evaluated (errors)
Gerd Bramerdorfer September 2 2018 Slide 75
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Speed Improvements for
Electric Machine and Drive
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 76
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 77
Many electrical engineers need to optimize electrical machines drives
Typically computationally expensive evaluation of designs isrequired
A lot of design parameters need to be varied (parameters of thegeometry control parameters different materials)
Engineers usually face a multi-objective problem(costs efficiency torque ripple)
Results should - as always - be available as fast as possible
This should serve as reference how speed improvements foroptimization runs for electrical machine designs can be achieved
a) from the theory of machine designsrsquo point of view
b) from an optimization strategy point of view
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
should determine the quality in terms of convergence and diversity
Gerd Bramerdorfer September 2 2018 Slide 71
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Gerd Bramerdorfer September 2 2018 Slide 72
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
should determine the quality in terms of convergence and diversity
Thus they are usually classified to
convergence quantifiers (eg epsilon measure general distance)
diversity quantifiers (eg spread metric generalized spread)
both convergence and diversity quantifiers (eg inversegenerational distance hypervolume indicator)
Gerd Bramerdorfer September 2 2018 Slide 73
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Gerd Bramerdorfer September 2 2018 Slide 74
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Other helpful measures
ldquoNew in Pareto frontrdquo Size of Pareto front
Distribution of Pareto variables(distinct optimal values)
Individuals per generation that cannot be evaluated (errors)
Gerd Bramerdorfer September 2 2018 Slide 75
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Speed Improvements for
Electric Machine and Drive
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 76
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 77
Many electrical engineers need to optimize electrical machines drives
Typically computationally expensive evaluation of designs isrequired
A lot of design parameters need to be varied (parameters of thegeometry control parameters different materials)
Engineers usually face a multi-objective problem(costs efficiency torque ripple)
Results should - as always - be available as fast as possible
This should serve as reference how speed improvements foroptimization runs for electrical machine designs can be achieved
a) from the theory of machine designsrsquo point of view
b) from an optimization strategy point of view
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Gerd Bramerdorfer September 2 2018 Slide 72
Source Lecture notes of Ciprian Zavoianu JKU
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
should determine the quality in terms of convergence and diversity
Thus they are usually classified to
convergence quantifiers (eg epsilon measure general distance)
diversity quantifiers (eg spread metric generalized spread)
both convergence and diversity quantifiers (eg inversegenerational distance hypervolume indicator)
Gerd Bramerdorfer September 2 2018 Slide 73
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Gerd Bramerdorfer September 2 2018 Slide 74
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Other helpful measures
ldquoNew in Pareto frontrdquo Size of Pareto front
Distribution of Pareto variables(distinct optimal values)
Individuals per generation that cannot be evaluated (errors)
Gerd Bramerdorfer September 2 2018 Slide 75
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Speed Improvements for
Electric Machine and Drive
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 76
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 77
Many electrical engineers need to optimize electrical machines drives
Typically computationally expensive evaluation of designs isrequired
A lot of design parameters need to be varied (parameters of thegeometry control parameters different materials)
Engineers usually face a multi-objective problem(costs efficiency torque ripple)
Results should - as always - be available as fast as possible
This should serve as reference how speed improvements foroptimization runs for electrical machine designs can be achieved
a) from the theory of machine designsrsquo point of view
b) from an optimization strategy point of view
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
should determine the quality in terms of convergence and diversity
Thus they are usually classified to
convergence quantifiers (eg epsilon measure general distance)
diversity quantifiers (eg spread metric generalized spread)
both convergence and diversity quantifiers (eg inversegenerational distance hypervolume indicator)
Gerd Bramerdorfer September 2 2018 Slide 73
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Gerd Bramerdorfer September 2 2018 Slide 74
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Other helpful measures
ldquoNew in Pareto frontrdquo Size of Pareto front
Distribution of Pareto variables(distinct optimal values)
Individuals per generation that cannot be evaluated (errors)
Gerd Bramerdorfer September 2 2018 Slide 75
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Speed Improvements for
Electric Machine and Drive
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 76
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 77
Many electrical engineers need to optimize electrical machines drives
Typically computationally expensive evaluation of designs isrequired
A lot of design parameters need to be varied (parameters of thegeometry control parameters different materials)
Engineers usually face a multi-objective problem(costs efficiency torque ripple)
Results should - as always - be available as fast as possible
This should serve as reference how speed improvements foroptimization runs for electrical machine designs can be achieved
a) from the theory of machine designsrsquo point of view
b) from an optimization strategy point of view
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Gerd Bramerdorfer September 2 2018 Slide 74
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Other helpful measures
ldquoNew in Pareto frontrdquo Size of Pareto front
Distribution of Pareto variables(distinct optimal values)
Individuals per generation that cannot be evaluated (errors)
Gerd Bramerdorfer September 2 2018 Slide 75
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Speed Improvements for
Electric Machine and Drive
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 76
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 77
Many electrical engineers need to optimize electrical machines drives
Typically computationally expensive evaluation of designs isrequired
A lot of design parameters need to be varied (parameters of thegeometry control parameters different materials)
Engineers usually face a multi-objective problem(costs efficiency torque ripple)
Results should - as always - be available as fast as possible
This should serve as reference how speed improvements foroptimization runs for electrical machine designs can be achieved
a) from the theory of machine designsrsquo point of view
b) from an optimization strategy point of view
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Quality measures
Hypervolume indicatorldquohow large is the area defined by the current Pareto-frontcompared to the previous one(s)rdquo
Other helpful measures
ldquoNew in Pareto frontrdquo Size of Pareto front
Distribution of Pareto variables(distinct optimal values)
Individuals per generation that cannot be evaluated (errors)
Gerd Bramerdorfer September 2 2018 Slide 75
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Speed Improvements for
Electric Machine and Drive
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 76
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 77
Many electrical engineers need to optimize electrical machines drives
Typically computationally expensive evaluation of designs isrequired
A lot of design parameters need to be varied (parameters of thegeometry control parameters different materials)
Engineers usually face a multi-objective problem(costs efficiency torque ripple)
Results should - as always - be available as fast as possible
This should serve as reference how speed improvements foroptimization runs for electrical machine designs can be achieved
a) from the theory of machine designsrsquo point of view
b) from an optimization strategy point of view
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Speed Improvements for
Electric Machine and Drive
Optimization Scenarios
Gerd Bramerdorfer September 2 2018 Slide 76
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 77
Many electrical engineers need to optimize electrical machines drives
Typically computationally expensive evaluation of designs isrequired
A lot of design parameters need to be varied (parameters of thegeometry control parameters different materials)
Engineers usually face a multi-objective problem(costs efficiency torque ripple)
Results should - as always - be available as fast as possible
This should serve as reference how speed improvements foroptimization runs for electrical machine designs can be achieved
a) from the theory of machine designsrsquo point of view
b) from an optimization strategy point of view
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Motivation
Gerd Bramerdorfer September 2 2018 Slide 77
Many electrical engineers need to optimize electrical machines drives
Typically computationally expensive evaluation of designs isrequired
A lot of design parameters need to be varied (parameters of thegeometry control parameters different materials)
Engineers usually face a multi-objective problem(costs efficiency torque ripple)
Results should - as always - be available as fast as possible
This should serve as reference how speed improvements foroptimization runs for electrical machine designs can be achieved
a) from the theory of machine designsrsquo point of view
b) from an optimization strategy point of view
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Categorization
Basic scope for action
Techniques for an efficient exploration of the design space
Advanced modeling techniques
Gerd Bramerdorfer September 2 2018 Slide 78
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 79
Basic scope for action
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 80
It is always recommended to considersymmetries of your machine design
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 81
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Symmetry considerations Modelsimplifications
Gerd Bramerdorfer September 2 2018 Slide 82
A1
A2
A3
B1
B2
B3
a1
b1
b2
a2
It is always recommended to considersymmetries of your machine design
As shown in [9] symmetrical threephase machines show periodicbehavior with regard to a sixth of anelectrical period
rArr If iron losses can be neglected orif iron losses of flux densitycharacteristics of different regionscan be extracted of the FE softwareand separately evaluated only asixth of an electrical period needs tobe analyzed
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 83
Techniques for an efficient
exploration of the design space
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 84
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
blank blankAnalysis of all individuals is done in parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 85
Inspired by evolution
A population evolves over time
Less individuals need to beanalyzed compared to grid search
Analysis of all individuals is donein parallel on a clusterrarr waiting for the slowest nodesleads to non-constant workload is not time-efficient
minus100 minus975 minus95 minus925 minus90 minus875 minus850
50
100
150
200
250
300
350
400
450
minusEfficiency
Co
sts
0 1 2 3 4 5 60
50
100
150
200
250
300
350
400
450
Torque ripple
Co
sts
Generation 100
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - generation-based approach
Gerd Bramerdorfer September 2 2018 Slide 86
Typical cluster work load
No continuous work load is achieved as the processneeds to wait for the slowest nodes the generatedlast design candidates for evaluation
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - asynchronous approach
Gerd Bramerdorfer September 2 2018 Slide 87
Every time when the analysis of anindividual is completed a newindividual is sent to the clusterrarr the cluster can be held at constantloadrarr less information is available onaverage when creating the i-thindividual compared to thegeneration-based approachrarr there is a trade offtests show that for optimizationswhere a considerable evaluationtime is spent for the analysis ofindividuals the convergence speedis highly increased [3]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Multi-stage analysis of particular machinedesigns
Gerd Bramerdorfer September 2 2018 Slide 88
Classifying individuals before computationally expensive calculations are
started - different possibilities for obtaining a rough estimate of the
performance for a possible design variant exist
Analytical forecast
Analysis of a reduced number of angularpositions (eg 0 and 15 el)
Analysis of a single current vector
Classification of a design by consideringresults for similar designs andperforming interpolation eg in [10]where Kriging was used
rarr If a promising design was found an automatizeddetailed analysis can be done afterwards
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 89
Advanced modelingtechniques
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the targets of an optimization scenario (ldquoMappingrdquo)
Gerd Bramerdorfer September 2 2018 Slide 90
After the initial phase asurrogate model isconstructed
The model is used forrunning the optimizationafterwards
In the end promisingdesigns are verified byFE-simulations
rarr Significant reduction of computation time can be achievedDetails can be found in [11]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 91
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
minus200 minus150 minus100 minus50 0 50 100
minus100
minus50
0
50
100
150
200
dminusaxis current in percent
qminus
axis
cu
rre
nt
in p
erc
en
t
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 92
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
A suitable interpolation function is used
(eg symbolic regression [9 12] or radial basis functions [2])
rarr Any other load point can be calculatedvery fast without using FE
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Conclusion
Gerd Bramerdorfer September 2 2018 Slide 93
Optimization of electric machines and drives is an emerging field
Knowledge of many different disciplines can help improving theperformance of designs and can significantly reduce the run time ofoptimizations
Even though computational power was significantly increasedapply engineering skills to appropriately define and perform youroptimization runs
Future aspects are eg with regard to
a ldquomore multi-physics and system-oriented optimizationrdquoincluding a tolerance analysis robustness evaluation to theoptimization scenario in order to avoid negative surprisesdriving cycle based evaluations of designs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Genetic algorithms - asynchronous approach
Gerd Bramerdorfer September 2 2018 Slide 87
Every time when the analysis of anindividual is completed a newindividual is sent to the clusterrarr the cluster can be held at constantloadrarr less information is available onaverage when creating the i-thindividual compared to thegeneration-based approachrarr there is a trade offtests show that for optimizationswhere a considerable evaluationtime is spent for the analysis ofindividuals the convergence speedis highly increased [3]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Multi-stage analysis of particular machinedesigns
Gerd Bramerdorfer September 2 2018 Slide 88
Classifying individuals before computationally expensive calculations are
started - different possibilities for obtaining a rough estimate of the
performance for a possible design variant exist
Analytical forecast
Analysis of a reduced number of angularpositions (eg 0 and 15 el)
Analysis of a single current vector
Classification of a design by consideringresults for similar designs andperforming interpolation eg in [10]where Kriging was used
rarr If a promising design was found an automatizeddetailed analysis can be done afterwards
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 89
Advanced modelingtechniques
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the targets of an optimization scenario (ldquoMappingrdquo)
Gerd Bramerdorfer September 2 2018 Slide 90
After the initial phase asurrogate model isconstructed
The model is used forrunning the optimizationafterwards
In the end promisingdesigns are verified byFE-simulations
rarr Significant reduction of computation time can be achievedDetails can be found in [11]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 91
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
minus200 minus150 minus100 minus50 0 50 100
minus100
minus50
0
50
100
150
200
dminusaxis current in percent
qminus
axis
cu
rre
nt
in p
erc
en
t
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 92
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
A suitable interpolation function is used
(eg symbolic regression [9 12] or radial basis functions [2])
rarr Any other load point can be calculatedvery fast without using FE
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Conclusion
Gerd Bramerdorfer September 2 2018 Slide 93
Optimization of electric machines and drives is an emerging field
Knowledge of many different disciplines can help improving theperformance of designs and can significantly reduce the run time ofoptimizations
Even though computational power was significantly increasedapply engineering skills to appropriately define and perform youroptimization runs
Future aspects are eg with regard to
a ldquomore multi-physics and system-oriented optimizationrdquoincluding a tolerance analysis robustness evaluation to theoptimization scenario in order to avoid negative surprisesdriving cycle based evaluations of designs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Multi-stage analysis of particular machinedesigns
Gerd Bramerdorfer September 2 2018 Slide 88
Classifying individuals before computationally expensive calculations are
started - different possibilities for obtaining a rough estimate of the
performance for a possible design variant exist
Analytical forecast
Analysis of a reduced number of angularpositions (eg 0 and 15 el)
Analysis of a single current vector
Classification of a design by consideringresults for similar designs andperforming interpolation eg in [10]where Kriging was used
rarr If a promising design was found an automatizeddetailed analysis can be done afterwards
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 89
Advanced modelingtechniques
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the targets of an optimization scenario (ldquoMappingrdquo)
Gerd Bramerdorfer September 2 2018 Slide 90
After the initial phase asurrogate model isconstructed
The model is used forrunning the optimizationafterwards
In the end promisingdesigns are verified byFE-simulations
rarr Significant reduction of computation time can be achievedDetails can be found in [11]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 91
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
minus200 minus150 minus100 minus50 0 50 100
minus100
minus50
0
50
100
150
200
dminusaxis current in percent
qminus
axis
cu
rre
nt
in p
erc
en
t
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 92
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
A suitable interpolation function is used
(eg symbolic regression [9 12] or radial basis functions [2])
rarr Any other load point can be calculatedvery fast without using FE
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Conclusion
Gerd Bramerdorfer September 2 2018 Slide 93
Optimization of electric machines and drives is an emerging field
Knowledge of many different disciplines can help improving theperformance of designs and can significantly reduce the run time ofoptimizations
Even though computational power was significantly increasedapply engineering skills to appropriately define and perform youroptimization runs
Future aspects are eg with regard to
a ldquomore multi-physics and system-oriented optimizationrdquoincluding a tolerance analysis robustness evaluation to theoptimization scenario in order to avoid negative surprisesdriving cycle based evaluations of designs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Gerd Bramerdorfer September 2 2018 Slide 89
Advanced modelingtechniques
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the targets of an optimization scenario (ldquoMappingrdquo)
Gerd Bramerdorfer September 2 2018 Slide 90
After the initial phase asurrogate model isconstructed
The model is used forrunning the optimizationafterwards
In the end promisingdesigns are verified byFE-simulations
rarr Significant reduction of computation time can be achievedDetails can be found in [11]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 91
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
minus200 minus150 minus100 minus50 0 50 100
minus100
minus50
0
50
100
150
200
dminusaxis current in percent
qminus
axis
cu
rre
nt
in p
erc
en
t
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 92
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
A suitable interpolation function is used
(eg symbolic regression [9 12] or radial basis functions [2])
rarr Any other load point can be calculatedvery fast without using FE
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Conclusion
Gerd Bramerdorfer September 2 2018 Slide 93
Optimization of electric machines and drives is an emerging field
Knowledge of many different disciplines can help improving theperformance of designs and can significantly reduce the run time ofoptimizations
Even though computational power was significantly increasedapply engineering skills to appropriately define and perform youroptimization runs
Future aspects are eg with regard to
a ldquomore multi-physics and system-oriented optimizationrdquoincluding a tolerance analysis robustness evaluation to theoptimization scenario in order to avoid negative surprisesdriving cycle based evaluations of designs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the targets of an optimization scenario (ldquoMappingrdquo)
Gerd Bramerdorfer September 2 2018 Slide 90
After the initial phase asurrogate model isconstructed
The model is used forrunning the optimizationafterwards
In the end promisingdesigns are verified byFE-simulations
rarr Significant reduction of computation time can be achievedDetails can be found in [11]
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 91
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
minus200 minus150 minus100 minus50 0 50 100
minus100
minus50
0
50
100
150
200
dminusaxis current in percent
qminus
axis
cu
rre
nt
in p
erc
en
t
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 92
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
A suitable interpolation function is used
(eg symbolic regression [9 12] or radial basis functions [2])
rarr Any other load point can be calculatedvery fast without using FE
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Conclusion
Gerd Bramerdorfer September 2 2018 Slide 93
Optimization of electric machines and drives is an emerging field
Knowledge of many different disciplines can help improving theperformance of designs and can significantly reduce the run time ofoptimizations
Even though computational power was significantly increasedapply engineering skills to appropriately define and perform youroptimization runs
Future aspects are eg with regard to
a ldquomore multi-physics and system-oriented optimizationrdquoincluding a tolerance analysis robustness evaluation to theoptimization scenario in order to avoid negative surprisesdriving cycle based evaluations of designs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 91
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
minus200 minus150 minus100 minus50 0 50 100
minus100
minus50
0
50
100
150
200
dminusaxis current in percent
qminus
axis
cu
rre
nt
in p
erc
en
t
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 92
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
A suitable interpolation function is used
(eg symbolic regression [9 12] or radial basis functions [2])
rarr Any other load point can be calculatedvery fast without using FE
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Conclusion
Gerd Bramerdorfer September 2 2018 Slide 93
Optimization of electric machines and drives is an emerging field
Knowledge of many different disciplines can help improving theperformance of designs and can significantly reduce the run time ofoptimizations
Even though computational power was significantly increasedapply engineering skills to appropriately define and perform youroptimization runs
Future aspects are eg with regard to
a ldquomore multi-physics and system-oriented optimizationrdquoincluding a tolerance analysis robustness evaluation to theoptimization scenario in order to avoid negative surprisesdriving cycle based evaluations of designs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
I-MotivationOutline II-IntroductionOptim Techniques III-Exemplary Scenario IV-AutomationSpeed Improvements
Modeling the characteristics of a particular machine design
Gerd Bramerdorfer September 2 2018 Slide 92
A certain number of currentvectors are defined based onthe max current density(SynRM) or the no loadresults (PMSM) eg using
iqnom = Tnm2
pz Ψpm
The current vectors areanalyzed using FE-simulations
A suitable interpolation function is used
(eg symbolic regression [9 12] or radial basis functions [2])
rarr Any other load point can be calculatedvery fast without using FE
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Conclusion
Gerd Bramerdorfer September 2 2018 Slide 93
Optimization of electric machines and drives is an emerging field
Knowledge of many different disciplines can help improving theperformance of designs and can significantly reduce the run time ofoptimizations
Even though computational power was significantly increasedapply engineering skills to appropriately define and perform youroptimization runs
Future aspects are eg with regard to
a ldquomore multi-physics and system-oriented optimizationrdquoincluding a tolerance analysis robustness evaluation to theoptimization scenario in order to avoid negative surprisesdriving cycle based evaluations of designs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Conclusion
Gerd Bramerdorfer September 2 2018 Slide 93
Optimization of electric machines and drives is an emerging field
Knowledge of many different disciplines can help improving theperformance of designs and can significantly reduce the run time ofoptimizations
Even though computational power was significantly increasedapply engineering skills to appropriately define and perform youroptimization runs
Future aspects are eg with regard to
a ldquomore multi-physics and system-oriented optimizationrdquoincluding a tolerance analysis robustness evaluation to theoptimization scenario in order to avoid negative surprisesdriving cycle based evaluations of designs
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Finally I am (we are) looking forwardto working with you in the field ofoptimization or
Gerd Bramerdorfer September 2 2018 Slide 94
Source httpshakanforsswordpresscom20140310are-you-too-busy-to-improve
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
Acknowledgment
Gerd Bramerdorfergerdbramerdorferjkuat
Department of Electrical Drivesand Power ElectronicsJohannes Kepler University LinzAltenbergerstraszlige 694040 Linz AUSTRIA
Thank you for your kind
attention
This work has been supported by the COMET-K2 ldquoCenter for SymbioticMechatronicsrdquo of the Linz Center of Mechatronics (LCM) funded by the Austrianfederal government and the federal state of Upper Austria and the participatingscientific and industrial partners The authors thank all supporters
I further want to thank Ciprian Zavoianu PhD for sharing his lecture notes for thiswork
The author has made reference to the work taken from other resources and for surethe rights are owned by the corresponding persons or organizationsThe only intention here is to teach attendeesin the presented subject(s)
Gerd Bramerdorfer September 2 2018 Slide 95
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
References
[1] G Bramerdorfer S Silber G Weidenholzer and W Amrhein ldquoComprehensive cost optimization study ofhigh-efficiency brushless synchronous machinesrdquo in 2013 IEEE International Electric Machines amp Drives Conference(IEMDC) 2013 pp 1126ndash1131
[2] G Weidenholzer S Silber G Jungmayr G Bramerdorfer H Grabner and W Amrhein ldquoA flux-based PMSM motormodel using RBF interpolation for time-stepping simulationsrdquo in 2013 IEEE International Electric Machines amp DrivesConference (IEMDC) 2013 pp 1418ndash1423
[3] A-C Zavoianu E Lughofer W Koppelstatter G Weidenholzer W Amrhein and E P Klement ldquoOn theperformance of master-slave parallelization methods for multi-objective evolutionary algorithmsrdquo in Proceedings of the12th International Conference on Artificial Intelligence and Soft Computing - ICAISC 2013 Springer Berlin Heidelberg 2013
[4] E Zitzler M Laumanns and L Thiele ldquoSPEA2 Improving the Strength Pareto Evolutionary Algorithm forMultiobjective Optimizationrdquo in Evolutionary Methods for Design Optimisation and Control with Application toIndustrial Problems (EUROGEN 2001) 2002 pp 95ndash100
[5] E Zitzler and L Thiele ldquoMultiobjective evolutionary algorithms a comparative case study and the strength paretoapproachrdquo IEEE Transactions on Evolutionary Computation vol 3 no 4 pp 257 ndash271 nov 1999
[6] FEMAGEnglish httpswwwprofemagch
[7] S Silber G Bramerdorfer A Dorninger A Fohler J Gerstmayr W Koppelstatter D Reischl G Weidenholzer andS Weitzhofer ldquoCoupled optimization in MagOptrdquo Proceedings of the Institution of Mechanical Engineers Part IJournal of Systems and Control Engineering 2015
Gerd Bramerdorfer September 2 2018 Slide 96
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Department of Electrical Drives and Power ElectronicsFaculty of Mechatronicshttpwwwealjkuat
[8] S Silber W Koppelstatter G Weidenholzer and G Bramerdorfer ldquoMagOpt - Optimization Tool for MechatronicComponentsrdquo in 14th International Symposium on Magnetic Bearings (ISMB14) 2014 pp 243ndash246
[9] G Bramerdorfer S Winkler M Kommenda G Weidenholzer S Silber G Kronberger M Affenzeller andW Amrhein ldquoUsing FE calculations and data-based system identification techniques to model the nonlinear behavior ofPMSMsrdquo IEEE Transactions on Industrial Electronics vol 61 no 11 pp 6454ndash6462 Nov 2014
[10] F Bittner and I Hahn ldquoKriging-assisted multi-objective particle swarm optimization of permanent magnet synchronousmachine for hybrid and electric carsrdquo in 2013 IEEE International Electric Machines and Drives Conference (IEMDC)May 2013 pp 15ndash22
[11] A-C Zavoianu G Bramerdorfer E Lughofer S Silber W Amrhein and E P Klement ldquoHybridization ofmulti-objective evolutionary algorithms and artificial neural networks for optimizing the performance of electrical drivesrdquoEngineering Applications of Artificial Intelligence vol 26 no 8 pp 1781ndash1794 2013
[12] G Bramerdorfer S M Winkler M Affenzeller and W Amrhein ldquoIdentification of a nonlinear PMSM model usingsymbolic regression and its application to current optimization scenariosrdquo in IECON 2014 - 40th Annual Conference ofthe IEEE Industrial Electronics Society 2014
Gerd Bramerdorfer September 2 2018 Slide 97
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Date of birth December 5 1988Nationality FinnishEmail ndash anttilehikoinenaaltofindash anttismeklabcom
Mini-biondash Thesis (2017) modelling AC winding losses in random-wound
machinesndash Post-doc at Aalto University (FIN) since Sep 2017ndash Author of the open-source SMEKlib FEA library for Matlabndash Sporadic consulting through SMEKlab Ltd
Eg powertrain optimization for an ebike hairpin winding design energystorage fault diagnostics
Actively involved in 3 startups Periodic services for others
DSc Antti Lehikoinen
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Contents - Antti
Winding lossesIron lossesndash Hysteresis skippedndash Classical eddies dampingndash Manufacture effects Inter-sheet currents Material degradation
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Letrsquos divide into three componentsndash ldquoDC lossesrdquo = ndash Eddy currents = uneven J within conductorndash Circulating currents = uneven division of total
current between parallel strands
Winding losses
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Phenomenon well-understoodNumerical analysis often required for accurate-enough predictionLong CPU timendash Complex geometries dense meshndash Short time-step Advanced models
Winding losses - Difficulties
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Much attention spent on eddy-currentsndash Squared-field derivativendash HomogenizationComparatively less on circulating currentsndash Circuit models (van der Geest)ndash Online FEA approaches (Lehikoinen)ndash Hybrid approaches (Roth)
Advanced winding loss models
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Widely-used post-processing approach
Idea sim Homogeneous B over a round wire assumed solve
(1D) E from times = Feedback from eddy-current density to B ignored
Losses from P sim int
Squared-field derivative (SFD)
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Prosndash No deep access to FEA solver neededndash Series of time-static simulations possible
parallelizationndash Easy to implementndash Decent accuracy oftenConsndash Accuracy issues with small skin-depthsndash Accuracy issues with large circulating currents
SFD pros and cons
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
ICEM18 paper by C Rothndash lsquoAnalytical Model for AC Loss Calculation Applied to Parallel
Conductors in Electrical Machinesrsquo
ndash Addresses some limitations of original SFDPoster session Wed (TT Thermal Losses and Efficiency Issues)
Recent advances to SFD
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Idea (harmonic analysis) replace fine winding structure by a complex reluctivityndash BH-loop lossesndash Reluctivity defined so that correct loss density
obtainedIdea (time-domain) the same only now is frequency-dependent ladder circuit approaches
Homogenization
Gyselinck Sabariego Dular lsquoTime-Domain Homogenization of Windings in 2-D Finite Element Models
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Different versions (semi-)indendently by Szucs Lehti and LehikoinenIdea eliminate winding domains by linear algebra magicndash Exact results compared to
standard FEAndash 10-100x speedupsndash All loss components
consideredPresentation (Wed TT2)lsquoA High-PerformanceOpen-Source Finite Element AnalysisLibrary for Magnetics in MATLABrsquo
Macro-element approaches
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Typical divisionndash Classical (macro-scale) eddy-current lossesndash Hysteresis lossesndash Excess lossesLots and lots of researchCovered todayndash Classical eddies in online computationsndash Manufacture effects Inter-laminar currents Material degradation
Iron losses
Difficult
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Caused by induced currents flowing withinlamination sheetsHigh frequency andor thick sheet dampingmay be significantndash B lags behind H BH loop widensndash Post-processing may not be accurateFirst-order 2D solution replace reluctivity by+
ndash For more info check eg Rasilo rsquo Finite-Element Modeling and Calorimetric Measurement of Core Losses in Frequency-Converter-Supplied Synchronous Machinesrsquo
Classical eddy current losses
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Galvanic contact between two laminations(usually in tooth area) closed loop if yokewelded circulating currentndash Often caused by burrs due to punchingndash Prone to thermal runaway core meltdown
Interlaminar currents
Picture (edited) Lee et al rsquoA Stator-Core Quality-Assessment Technique for Inverter-Fed Induction Machinesrsquo
Modellingndash 3D but decent solution
with 2D boundary layermodel
ndash Random positions of contacts problematic
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Punching and laser-cuttingndash Decrease permeability especially around 08-12 Tndash Increase hysteresis losses
Can be identified by eg using different sample widthsin Epstein frame
Material degradation
BH and core loss measurements from Sundaria et al rsquoHigher-Order Finite Element Modeling of Material Degradation Due to Cuttingrsquo
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Typical approachndash Layers near cut edgesndash Individual BH curve for
each layer+ Software independent- Tedious dense meshNovel approach by Sundariandash Continuous mat model= ( )(1 minus )
ndash Good performance with higher-order elementsPresentation (Tue TT4) rsquoLoss Model for The Effects of Steel Cutting in Electrical Machinesrsquo
Modelling material degradation
From Rasilo et al rsquo Analysis of Iron Losses on the Cutting Edges of Induction Motor Core Laminationsrsquo
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-
Advanced loss computation approaches needed tondash Shorten computation timesndash Include non-standard (yet significant) loss components Some presented here at ICEM
Some access to software usually needed
Thanks for coming
anttismeklabcom
Conclusion
THANK YOU
Gerd and Antti
- Appendix
-