MODERN OPTIMIZATION OF ROTATING ELECTRIC MACHINES · Place of Birth: V¨ocklabruck, Austria ......
114
MODERN OPTIMIZATION OF ROTATING ELECTRIC MACHINES Gerd Bramerdorfer Department of Electrical Drives and Power Electronics, Johannes Kepler University Linz, Linz, Austria [email protected]Antti Lehikoinen Department of Electrical Engineering Aaalto University [email protected]
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
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
