Sunteți pe pagina 1din 38

Identification for industrial model-

based control
Vidar Alstad
Department of Chemical Engineering
NTNU

June 8, 2005

1
NOT IN ORIGINAL PRESENTATION!
Title: "Identification for industrial model-based control. "
Deatils: The committee would like a survey of what
kind of models (soft/hard) are used by industry for
model predictive control (MPC), how the models are
identified, and the expected future direction. The
candidate should also address the following question:
"Why does industry today use almost exclusively black-
box (soft) models?"

2
Outline

Scope of presentation
Introduction & brief history of MPC
Model identification
Linear models used in industrial MPC
Non-linear models in industrial MPC
Future directions
Why still black-box models?
Concluding remarks

3
Scope of presentation
Model based control (Brosilow and Joseph, 2001 )

Control systems that explicitely embed a


process model in the control algorithm

Examples:
Internal model control (Morari and Zafiriou, 1989)
Model predictive control (Richalet et al, 1976, Cutler and Ramaker,
1979)
Black-box and first principles

Black-box/Soft Hard
Structure Empirical/data-based First principles

Mathematical Linear/non-linear Linear/non-linear


form
4
Scope of presentation
Types of models formulations used in MPC
Predictive models
Based on the current measured variables and the current and
future inputs, the model must predict the future outputs
(Rossitier, 2003)

Steady-state models
Often steady-version of the above dynamic model or a separate
comprehensive model (Qin and Badgwell, 2003)

Continuous processes

5
Model predictive control
(Gorinevsky, 2005)

control horizon

1. At each time step, compute the 2. Apply the first value of the
optimal control inputs over the computed control input into
control horizon by solving an open the process
loop optimization problem over
the prediction horizon taking 3. At the next time step, redo
constraints into account the calculation
6
Industrial MPC formulation

Prediction horizon Control horizon

Prediction model

7
Brief history of industrial MPC
LQR (Kalman, 1964)
Little impact in process industries due to lack of constraint handling
(Richalet et al., 1976, Garcia et al. 1989)
IDCOM and DMC (Richalet et al., 1976, Cutler and Ramaker, 1979)
Input/output representation
Ad hoc constraint handling
QDMC (Cutler et al., 1983)
Explicit constraint handling (improved DMC)
IDCOM-M, HIECON, SMCA and SMOC (late 1980-1990s)
State space representation
Hard constraints and priorities
DMC-plus and RMPCT (2000+)
Improved identification technology
Non-linear MPC (Aspen, DOT-products)
8
Industrial usage of MPC
Factors for widespread usage in process industries
(Piche, et al. 2000)
Open-loop settling time minutes or hours
Well suited for multivariable control (MIMO)
Explicit constraint handling
Empirical modelling tools
Standard in refining, chemical, petrochemical, pulp and
paper and food processing.
Many industrial commercial vendors.

9
Non-exhaustive MPC Vendor List
(Allgwer, 2004)

ABB Trieber Controls Objectspace


ACT Yokogawa APC Optimal Control Research
Adaptics US Process Control L.L.C.
Pavilion Technologies
Adaptive Resources Eldridge Engineering Inc.
Adersa Predictive Control Ltd.
Elsag Bailey
Aspen Technology Predictor
Aurel Systems Inc. Envision Systems Inc.
Process System Consultants
Batch CAD Gensym
Bonner and Moore RSI
Enterprise Control Technologies
Brainwave Simulation and Advanced
Fantoft Process Group
C.F. Picou and Associates Simtech
Chemstations MATHWORKS
Texas Controls Inc.
Comdale Technologies Honeywell
Control Arts Inc. Inferential Control Company
Control Consulting Inc.
IntellOpt
Control Dynamics
Controlsoft Incorporated Knowledgescape
Cybosoft MDC Technology
Cybernetica
Neuralware
DOT Products
Nexus Engineering

10
Identification for MPC
Model identification is clearly the In a typical commissioning project,
Achilles heel of MPC (or any modeling efforts can take up to 90%
other model-based controller of the cost and time (Andersen and
design technique) (Ogunnaike and Kummel, 1992).
Ray, 1994)

Basics of identification

Introduce a sequence of inputs ujk


PROSESS
yjk contain the process information
By treating the time series of yk with
Identification uk , a model can be estimated.
method
Near all model types need
experiments
Model
11
Steps in industrial identification
(Sderstrm and Stoica, 1989)

Instrument verification
Pre-test Time to steady-state (TSS)
Data for initial identification

Sequential
Test protocol
Simultanious
Signal type (PRBS or step)

Model Model formulation to use

Model identification Model parameter


estimation
Equation error
Output error
Model validation
Model validation: Measure
goodness of model
No
12 Accept?
Pre-test and test protocol
Pre-test Pre-testing (Seborg et al., 2004)
Estimate process gains
Time constants
Test protocol
Time delays
Plant instrumentation verification
Model
Test protocol
Signal type
Model identification
Sequential vs. simultaneous
Closed loop or open loop
Model validation

Accept?

13
Test signals for MPC identification
Data from tests should be informative, e.g. contain information
on sufficient distinct frequencies. (Ljung, 1999).
MPC relevant signal design (McLellan, 2004)

Require good estimate on


power
steady-state gain and
slower dynamics
Higher frequencies
handled by the regulatory
control layer.

frequency

Industrial test signals (McLellan, 2004)


Step signals
14 Pseudo-Random Binary signals (PRBS)
Test signals for MPC identification
STEP PRBS
u y u y

t t t t

Provides very good information Provides good information on steady-state


on steady-state gain gain and higher frequencies.

power power
Limited information on
higher frequencies

frequency frequency
Input magnitude Input magnitude
Duriation Duration
Minimum switching time
15
Desired frequency content
Test signals and identification
( Qin and Badgwell, 2003; Conner and Seborg, 2004; Morari and Lee,
1999; Li and Lee, 1996)
Open loop and sequential SISO identification
Step Easy identify model structure
No information on directionality
Easy to administer and
easier to interpret data Good individual SISO match (either
Long test time step or frequency response), poor
MIMO model.
MISO identification
Simultanious excitation signals
Process upset
Output models fit one-by-one
Open loop and simultanious Better disturbance model
No information on directionality

16
Model formulation in industrial MPC
(Qin and Badgwell, 2003)

Non-linear
Pre-test Input/Output models
Nonlinear
Nonlinear neural net (NNN) +
auto-regressive with exogenous state space
inputs (ARX) Hybrid models
Test protocol State space
Nonlinear state space model
(LSS) (NNN)

Model

Empirical First principles


Model identification

Input/Output models
Finite impulse response (FIR)
Model validation Finite step response (FSR)
Laplace transfer function (TF)
Auto-regressive with exogenous
Accept? inputs (ARX)
State space
Linear state space model (LSS)
Linear
17
Non-linear

Linear models
Empirical
Majority of industrial MPC
applications use linear empirical
models (Qin and Badgewell, Linear

2003)

Input/Output models
Finite impulse response (FIR)
Finite step response (FSR)
Laplace transfer function (TF)
Auto-regressive with exogenous inputs
(ARX)
State space
Linear state space model (LSS)
18
Non-linear

Input/output models
FIR and FSR Empirical
First principles
Nonparametric models
Finite Impulse Response Linear

Parameters hi are the sampled


output after a unit impulse input
Finite Step Response

Parameters si is the sampled


output after a unit step input

19
Input/output models Non-linear

FIR and FSR


(Qin and Badgwell, 2003; Ljung, 1999) Empirical
First principles

General Finite Impulse Response (FIR)


Linear

General Finite Step Response (FSR)

20
Input/output models Non-linear

FIR and FSR


(Zhu et al., 2000, Ljung, 1999; Qin and Badgwell, 2003) Empirical
First principles

Linear

Finite impulse response (FIR) and Finite step response (FSR) model
+ Can handle complex dynamics
Time delays, inverse response
+ Little prior process knowledge needed
Time to steady state (TTSS)
Nu,Nd,Nv
+ Model identification
+ No bias in parameters due to measurement noise
Bias due to truncation
Over-parametrized
Sample time selected so 30-120 coefficients describe the full open loop
response
Cannot handle unstable or integrating processes
21
Output feedback
Input/output models Non-linear

ARX
(Zhu et al., 2000; Ljung, 1999; Qin and Badgwell, 2003) Empirical
First principles

Parametric models
Linear

Autoregressive model with exogenous inputs (ARX)


Example

General MIMO

Polynomial matrices, e.g.

22
Input/output models Non-linear

ARX
(Zhu et al., 2000; Ljung, 1999; Qin and Badgwell, 2003) Empirical
First principles

Same autoregressive part for inputs and


disturbances Linear

Disturbances enter near input of v


process (same denominator term)
Ev
Handles stable, unstable and integrating
dynamics
u
Model order B S S A-1
y
Other parametric models used
Ed
Box-Jenkins
Transfer function d
(converted to discrete time)

23 FIR special type of ARX


Non-linear

State space models


(Qin and Badgwell, 2003)
Empirical
First principles

Discrete state space


Linear

Handles stable, unstable and integrating processes


Systematic output feedback (Kalman filter)
Distinction between controlled and feedback variables
Unmeasured disturbance models
Linearized first principle model
Used in research literature

24
Linear model identification
(Qin and Badgwell, 2003)

Prediction error methods


Pre-test
Minimize a least square criterion

Test protocol

using either
Model
1. Equation error approach (one step ahead prediction)
2. Output error approach (multi step ahead prediction)
Model identification
Finite impulse and step response (FIR and FSR)
yield linear least square.
Model validation

Accept?
Predictor

25
Linear model identification ARX
(Qin and Badgwell, 2003)

Prediction error

Equation error identification approach Output error identification approach


Linear least square Nonlinear parameter estimation

Past measurements fed back in Past estimates fed back in model


model Numerically more challenging
Preferred industrial implementation Gauss-Newton methods
Gradient descent methods
Parameter estimates biased given
white measurement noise Global methods
Parameter estimates unbiased
One-step ahead prediction
26 Multi-step ahead prediction
Linear model identification
state space models
Based on subspace model identification (Van Overschee, 1996)
Efficient method for estimating MIMO models as compared to
existing PEM methods (Favoreel, et al, 2000)

Two step procedure


1. Estimate the state sequence from input/output data
2. Linear regression to find the system matrices

Can yield suboptimal estimates as compared to existing PEM


methods (Van Overschee, 1996)

27
Linear models Summary
(Qin and Badgwell,2003; Zhu and Butoyi, 2002; Ljung, 1999)

Nonparametric models Parametric models


(FIR and FSR) (ARX and state space)
Bias error due to truncation - Order selection (ARX)
Poor with fast/slow dynamics Ad hoc bias sceeme output
Non-compact. feedback (ARX)
Unable to model unstable and
integrating processes
Ad hoc bias sceeme output + No bias (if sufficient order)
feedback (ARX)
+ Compact
+ Handles unstable and
integrating processes
+ Linear least square for + MIMO identification (state
parameter estimation space)
+ Little process knowledge + Output feedback (state space)
needed
28
Model formulation in industrial MPC
(Qin and Badgwell, 2003)

Non-linear
Pre-test Input/Output models
Nonlinear
Nonlinear neural net (NNN) +
auto-regressive with exogenous state space
inputs (ARX) Hybrid models
Test protocol State space
Nonlinear state space model
(LSS) (NNN)

Model

Empirical First principles


Model identification

Input/Output models
Finite impulse response (FIR)
Model validation Finite step response (FSR)
Laplace transfer function (TF)
Auto-regressive with exogenous
Accept? inputs (ARX)
State space
Linear state space model (LSS)
Linear
29
Non-linear

Non-linear
empirical models Empirical
First principles

Linear

Input/Output models
Nonlinear neural net (NNN) + auto-
regressive with exogenous inputs (ARX)
State space
Nonlinear state space model (LSS) (NNN)

30
Non-linear

Input/output
(Piche et al., 2000, Qin and Badgwell, 2003)
Empirical
Pavilion First principles

Steady-state nonlinear model superimposed on


a linear dynamic model Linear

Second order linear model Steady state nonlinear part modeled


for dynamics as a bounded derivative neural net.
Open loop step response Trained using historical closed loop
for identification data

31
Non-linear

State space
( Zhao et al. ,1998;2001)
Empirical
First principles
Aspen Apollo
Linear

Claimed that above model can approximate any discrete time


nonlinear process with fading memory (Sentoni et al., 1998)
Bounded derivative network (Turner and Guiver, 2005)
Calculated bounds (max/min) on the gains
Globaly constrained, i.e. smooth transition to a linear approximation
(constant gain) in regions of extrapolation.
Identification
MISO identification of linear state space model using Principal
Component Analysis and Partial least squares methods
Neural net trained on prediction error for the linear model
32
Non-linear

Non-linear first principle


models Empirical
First principles

Mass, energy and impulse conservation


equations (Kouvaritakis and Cannon, 2001) Linear

Empirical data needed


Plant data not sufficient
Uncertain parameters estimated using least square (Young, et al.
2001)
Robustness and reliablility of identification algorithm
Art in estimation of the parameters. (Young, et al. 2001)
Nonlinear test signal design
33 Open topic (Morari et al., 1999)
Future directions
Linear models
Parametric models (Qin, 2005)
State space representation (Qin and Badgwell, 2003)
Subspace identification methods
Output feedback
Non-linear models
Non-linear empirical models (Piche et al., 2000)
Extensive testing
First principle models
Modeling and identification
Servo control problem with changes in operating point

Identification
MIMO identification
34
Why does industry still use black-box models?

Historical reasons Why not first principle nonlinear


Regulator problem models?
Sufficient performance using a Modeling and maintenance cost
(Pische, 2000)
linear models (Keep it simple!)
Justification criteria
Non-linearities handled by
variable transformations Often process and operation
specific
Easy to identify
Test signal and identification
Linear least square (ARX, (Morari and Lee, 1999)
FIR/FSR)
Modeling tools
Subspace methods
Non-convex online optimization,
Efficient online optimization computationally expensive (Morari
and Lee, 1999)
Empircal nonlinear models
next step beyond linear
modeling of chemical
processes. (Henson and
35 Seborg, 1997)
Concluding remarks
Predictive models for MPC
Linear empirical models
Sequential and open loop testing
Non-linearity addressed using empirical models (e.g. neural net)
First principle models not widely used due to cost of design and
maintenance
Improvements in identification technology would have an
positive impact on MPC technology

Acknowledgements:
Bjrn Glemmestad, Vinay Kariwala, Audun Faanes, Tor Steinar Schei,
Stig Strand, Kjetil Fjaalestad, Svein Olav Hauge and Olav
Slupphaug

36
Bibliography
1. C. Brosilow and B. Joseph. Techniques of Model-Based Control. Prentice Hall International Series, 2001.
2. M. Morari and E. Zafiriou, Robust Process Control, Prentice-Hall, Englewood Cliffs, 1989
3. J. A. Rossitier, Model based predictive control A practical Approach, CRC Press, 2003
4. J. R. Parrish and C. B. Brosilow, Inferential Control Algorithms, Automatica, 21(5):527-538, 1985.
5. J. Richalet, A. Rault, J.L. Testrud, J. Papon, Algorithmic control of industrial processes, In Proceedings of
the 4th IFAC symposium on identification asd system parameter estimation ., 1976.
6. C.R. Cutler and B.L. Ramaker. Dynamic matrix control- a computer control algorithm. AIChE national
meeting, Houstin TX, 1979
7. R. Kalman, When is a linear control system optimal?, Journal of Basic Engineering, Transactions on
ASME, Series D, 86
8. B. Kouvaritakis and M. Cannon (Editors), Nonlinear predictive control theory and practice, IEE Control
engineering series 61, 2001.
9. M. Nikolaou, Model Predictive Controllers: A Critical Synthesis of Theory and Industrial Needs. accepted,
Advances in Chemical Engineering Series, Academic Press, 1998
10. S. Piche, B. Sayyar-Rodsari, D. Johnson and M. Gerules, Nonlinear model predictive control using neural
networks. IEEE Control Systems Magazine, 2000.
11. C.R. Cutler, A. Morshedi and J. Haydel. An industrial perspective on advanced control. In AIChE annual
meeting, Washington, DC, 1983.
12. E. Demoro, C. Axelrud, D. Johnston and G. Martin, Neural network modelling and control of polypropylene
process, Society of Plastics Engineers International Conference, Houston TX, 1997.
13. B.A. Ogunnaike and W.H. Ray. Process Dynamics, Modeling and Control. Oxford University Press, 1994
14. Y. Zhu, E. Arrieta, F. Butoyi, and F. Cortes. Parametric versus nonparametric models in MPC process
identification. Hydrocarbon Processing, 79(2), 2000.
15. T. Sderstrm and P. Stoica, System identification, Prentice Hall Int., 1989
16. D.E. Seborg, T. Edgar and D.A. Mellichamp, Process Dynamics and Control, 2nd edition, Wiley, 2004.
17. H.W. Andersen, and M. Kummel, Evaluating estimation of gain directionalityPart 2: a case study of binary
distillation. Journal of Process Control, 2 (2), 6786, 1992
18. W. Li, and J.H. Lee, Frequency-domain closed-loop identification of multivariable systems for feedback
control. AIChE Journal, 20, 28132827, 1996
19. W. S. Levine, The Control Handbook, CRC Press, 1995
37 20. L. Ljung, System Identification, 2nd ed, Prentice Hall, 1999
21. Y. Zhu, Case studies on closed loop identification for MPC, Control Engineering Practice, 10, 2002
Bibliography (2)
22. Y. Zhu, Multivariable process identification for MPC, Journal of Process Control, 8(2), 1998
23. J.S. Conner and D.E Seborg, An evaluation of MIMO Input design for Process identification,
Ind.Eng.Chem.Res, 43, 2004
24. Y. Zhu, E. Arrieta, F. Butoyi and F. Cortes, Parametric versus nonparametric models in MPC process
identification,
25. H.Zhao, J. Guiver and G. Sentino, An identification approach to nonlinear state space model for industrial
multivariable model predictive control, Proceedings of the American Control Conference, June, 1998
26. H. Zhao, J. Guiver, R. Neelakantan, L.T. Biegler, A nonlinear industrial model predictive controller using
integrated PLS and neural net state-space model, Control Engineering Practice, 9, 2001
27. R.B. Gopaluni, R.S. Patwardhan and Shah, S.L., MPC relevant identification tuning the noise model,
Journal of Process Control, 14, 2004
28. P., Van Overschee, B, DeMoor, Subspace Identification for Linear Systems: Theory-Implementation-
Applications. Kluwer Academic Publishers 1996.
29. G. B. Sentoni, L. T. Biegler, J.B. Guiver, & H. Zhao, State space non-linear process modeling, AiChE
Journal, 1998.
30. P. Turner and J. Guiver, Introducing the bounded derivative network-superceding the application of neural
networks in control, Journal of Process Control, 15, 2005.
31. Aspen Tech., Aspen Apollo - Non-Linear Advanced Control Solution for Polymers,
http://www.aspentech.com/brochures/Aspen_Apollo.pdf, 2005
32. F.Allgwer, Presentation held at APACT04, Bath, April 26-28, 2004
33. S.J. Qin and T.A. Badgwell, An overview of nonlinear model predictive control applications, In J.C. Kantor,
C.E. Garciea & B. Carnahan (Eds.), Nonlinear model predictive control. Basel: Birkhauser, 2000
34. M. McLellan, Course notes, CHEE825/435, 2004
35. N. Bonavita, R. Martini and T. Matsko, Improvment in the Performance of Online Control Applications via
Enhanced Modeling Techniques, ERTC Computing, Milano, 2003
36. E.F. Camacho and C. Bordons, Model Predictive Control, Springer, 2000
37. R.E. Young, R.D. Burtusiak and R.W. Fontaine, Evolution of an Industrial Nonlinear Model Predictive
Controller, CPC 6, Arizona, 2001
38. M. Henson and D.E. Seborg, Nonlinear Process Control, Prentice Hall, 1997

38

S-ar putea să vă placă și