COMOD 2014 St. Petersburg, Russia, 2-4 July 2014 Smart Adaptive Methods in Modelling and Simulation...

30
COMOD 2014 St. Petersburg, Russia, 2-4 July 2014 Smart Adaptive Methods in Modelling and Simulation of Complex Systems Esko Juuso Control Engineering Group, Faculty of Technology University of Oulu

Transcript of COMOD 2014 St. Petersburg, Russia, 2-4 July 2014 Smart Adaptive Methods in Modelling and Simulation...

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Smart Adaptive Methods in Modelling and Simulation

of Complex SystemsEsko Juuso

Control Engineering Group,

Faculty of Technology

University of Oulu

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

EUROSIM Federation of European Simulation

Societies OULU

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

EUROSIM Federation of European Simulation

Societies

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Detection of operating conditions- system adaptation

-fault diagnosis, condition monitoring, quality

Dynamic simulation- controller design, prediction

Intelligent analysers-sensor fusion

-software sensors-trends

Intelligent control-adaptation

-model-based

Measurements-on-line analysers

-DSP

Intelligent actuators- model-based

Control Engineering Group

Competence Pyramid

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Outline• Background

– Soft computing: fuzzy set systems– Hard computing: statistical analysis

• Modelling & Simulation– Data + Knowledge + Decomposition

• Linguistic equation (LE) systems– Generalised moments and norms– Nonlinear scaling– Genetic tuning

• Application examples• Conclusions

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Symptom generation

-limit values, parameter esimates-analytic, heuristic

-condition monitoring-statistical process control (SPC)

Nonlinear process control

- feedback-fuzzy, neural, sliding mode

- adaptation (on-line, predefined)- model-based (FF, IMC, MPC)

- high-level

Soft sensors

-data-collection-pre-processing

-normalisation and scaling-interpolation

-data quality, outliers-signal processing-feature extraction

-sensor fusion

Nonlinear multivariable methodologies

- steady-state & dynamic-decomposition, clustering, composite models

-mixed models-development and tuning

-statistical, fuzzy, neural, genetic

Classification and reasoning methodologies

-rule-based, fuzzy, neural, support vector-artificial immune systems

-qualitative models, search strategies

Classification and reasoning

-case-based reasoning (CBR), models-fault and event trees

-cause-effect relationships-novelty detection

Detection of operating conditions

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Steady-state modelling: Data

Statistical analysis• Interactions

– Linear, quadratic & interactive Response surface methodology (RMS)

• Reduce dimensions– Principal component

analysis (PCA)– Partial least squares

regression (PLS)

Artificial neural networks

• Linear networks– Regression– Recursive tuning

• Multilayer perceptron– Nonlinear activation

• Learning– Backpropagation– Advanced optimisation

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Steady-state modelling: Knowledge

Fuzzy arithmetics• Extension principle• Interval arithmetics• Horizontal systems

Rules and relations• Linguistic fuzzy• Takagi-Sugeno fuzzy• Singleton• Fuzzy relational

modelsType-2 fuzzy sets• Uncertainty about the

membership functions

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Fuzzification

Fuzzyreasoning

Fuzzyrulebase

Defuzzification

Fuzzy

Crisp

Fuzzy

Crisp

Fuzzy

Fuzzy relations

Fuzzy set systems

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Steady-state modelling: Decomposition

Modelling

• Subprocesses• Hierachical• Composite models

– Linear parameter varying (LPV)

– Piecewise affine (PWA)

– TS fuzzy models– Ensemble of

redundant neural networks

Clustering

• Hierarchical• Partitioning: K-means• Fuzzy

– Fuzzy c-means (FCM)– Subtractive

• Neural: SOM

• Shape (Gustafson-Kessel)

• Robust• Optimal number

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Complex applications: Fuzzy set systems

Domain expertise

Datamining

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Expert Systems+ Extracting expert knowledge

- Complexity- Handling of uncertainty- Testing

Fuzzy Set Systems+ Handling of uncerainty+ Natural compromises+ Easy to build (small systems)+ Explanations - Tuning (complex systems)- (Doubts about stability)

Linguistic Equations+ Very compact+ Combining knowledge+ Generalisation+ Adaptive tuning+ Easier testing

- Structure Restrictions

Genetic Algorithms+ Large search space+ Global/local optimisation+ Design

- Computer Time Consuming- Not for Control (off-line)

Neural Networks+ ”Automatic” Modelling+ Black Box Modelling+ Precision (small systems)- Only for Fragments- Explanations- Safety- Precision (complex systems)

EXPERTISE

DATA

Neuro-fuzzyNN Structures

Knowledge-basealternatives

Rules

Chaos Theory•Risk Analysis•Economical factors

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Fuzzy set systems Linguistic equation systems

Meaning

Linear interactions

Smart adaptiveapplications- Modelling- Control- Diagnostics

How to define??

Hard computing??

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Linguistic relations- Selected and scaled data

Data Data selection- Outliers- Suspicious

Nonlinear scaling- Feasible ranges- Membership definitions- Membership functions

Adaptation of scaling functions- Generalised norms and moments- Constraints- Case specific

Variable grouping- 3-5 variables- Include/exclude- Correlation- Causality

Selected variable groups

Domain expertise

Linguistic equation alternatives- Linear regression- Case specific

Adaptation- Manual- Neural- Genetic

Selected equations Final variable groups

Manually defined equations

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Statistical analysis: norms• A generalised norm about the origin

which is the lp norm

• Special cases

– absolute mean

– rms value

• Positive and negative values

sNN

,)1

()( /1

1

)(/1 pN

i

p

ipp

p

p xN

MM

.)(

pp

p xM

,1

1

)()(

1

)(

N

iiav x

Nxx

,)1

( 2/1

1

2)()(

2

)(

N

iirms x

Nxx

p is a real number

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Generalised norms• equal sized sub-blocks

• A maximum from several samples

• Increasing

,)(1

)(1

/1

1

/1

1

/1

pK

ii

p

S

pK

i

ppi

p

Sp

pKSS

S MK

MK

M

pi

p

Ki

p MMS

/1

,...,1)(max)max(

qqpp MM /1/1 )()(

qp

,1

1)(

1

)(

N

i ix

Nx

2/1

1

2)(

2

)( )1

(

N

iix

Nx ,

1

1

)(

1

)(

N

iix

Nx

… …

Recursive analysis!

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Generalised moments• Normalised moments

• Skewness– Positive– Symmetric– Negative

• Generalised moment

• Locally linear if possible• Corrections for corner points• Core • Support

k = 3 Skewnessk = 4 Kurtosis

kX

k

k

XEXE

)(

03

03 03

k

X

k

p

p

k

MXE

)(

Central value

)](,)[( hjl cc

)]max(),[min( jj xx

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

LE: nonlinear scaling linear models (interactions)

Data

Meaning

Expertise

Knowledge-based information: labels to numbers

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Second order polynomialsTuning

(1) Core

(2) Ratios

(3) Support

• Centre point

• Corner points

• Calculation

)max(,)(,)(),min( jjhjlj xccx

jc

)](,)[( hjl cc

3,3

1j

)]max(),[min( jj xx

3,3

1j

jjj

jjj

jjj

jjj

cb

ca

cb

ca

)3(2

1

,)1(2

1

,)3(2

1

,)1(2

1

)min(2

)min(22

)(4

)max(22

)(4

)max(2

2

2

jj

jjjj

jjjjj

jjjj

jjjjj

jj

j

xxwith

cxxwitha

xcabb

xxcwitha

xcabb

xxwith

X

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

LE models: Dynamic simulator

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Genetic tuning

• Membership definitions– Parameters

– No penalties

• Normalised interactions

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Lagphase

Exp.phase

Steadystate

Decision system

X

X

X +

Integration

Prediction

Fuzzy weighting

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Submodels

CO2 forecast

OTR forecast

DO forecast

Measurements

Volumetric mass transferCoefficient, kLa

Fuzzy LE blocks

Note: 3 phases & 3 models / phase 9 interactive dynamic models!

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

LE Application examples: Control

• Energy: – Solar power plant

• Environment: – Water circulation & wastewater treatment

• Pulp&Paper: – Lime kilns

Length > 100 mSlow rotation: rotation time 42-45 s

~ 4 m

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

• Setpoint tracking

• Cloudy conditions

• Optimisation

Solar thermal power plantSolar thermal power plant

www.psa.es

Principle: lower irradiation lower temperaturesOperator can choose the risk level: smooth … fast

Clouds High temperature are risky Cloudy conditions are detected from fluctuations of irradiation Working point is limited Further limitations for the setpoint

Constrained optimisation:-Temperature (< 300 oC)

- Temperature increase (< 90 oC)

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Solar thermal power plant• Intelligent control

– Adaptation, braking, asymmetrical action– Automatic smart actions– Disturbances are handled well if the

working point is on a good level

• Intelligent indices– react well to disturbances (clouds, load,

…)

• Model-based limits for the working point Better adaptationSmooth adjustable operation A good basis for optimised operation within a Smart Grid

MODEL-BASEDCONTROL

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

LE Application examples: Diagnostics

• Stress indices– Cavitation

• Condition indices– Lime kiln

• Fatigue

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

Conclusions

• Soft computing– Expertise– Fuzzy reasoning

• Hard computing– Data– Statistical analysis

• Generalised norms and moments

Complex systems• Interactions

– Fuzzy set systems– Linguistic equations

• Meaning– Membership definitions Membership functions

• Nonlinear scaling

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

EUROSIM Federation of European Simulation

Societies

34th Board Meeting in Vienna, February 2012,

NSS became an observer member of EUROSIM

COMOD 2014St. Petersburg, Russia, 2-4 July 2014

EUROSIM 2016September 13-16, 2016, Oulu,

Finland

The 9th EUROSIM Congress on Modelling and Simulation

Oulu City Theatre

30