1

Self-optimizing control: Simple implementation ofoptimal operation

Sigurd Skogestad

Department of Chemical EngineeringNorwegian University of Science and Tecnology (NTNU)Trondheim, Norway

Benelux Control Meeting, March 2008

PART 2: Effective Implementation of optimal operation using Off-Line Computations

2

Outline

• Implementation of optimal operation

• Paradigm 1: On-line optimizing control

• Paradigm 2: "Self-optimizing" control schemes– Precomputed (off-line) solution

• Control of optimal measurement combinations– Nullspace method

– Exact local method

• Link to optimal control / Explicit MPC

• Current research issues

3

Optimal operation

• A typical dynamic optimization problem

• Implementation: “Open-loop” solutions not robust to disturbances or model errors

• Want to introduce feedback

4

Implementation of optimal operation

• Paradigm 1: On-line optimizing control where measurements are used to update model and states

• Paradigm 2: “Self-optimizing” control scheme found by exploiting properties of the solution in control structure design– Usually feedback solutions

– Use off-line analysis/optimization to find “properties of the solution”

– “self-optimizing ” = “inherent optimal operation”

5

Implementation: Paradigm 1

• Paradigm 1: Online optimizing control

• Measurements are primarily used to update the model

• The optimization problem is resolved online to compute new inputs.

• Example: Conventional MPC

• This is the “obvious” approach (for someone who does not know control)

6

Example paradigm 1: Marathon runner

• Even getting a reasonable model requires > 10 PhD’s … and the model has to be fitted to each individual….

• Clearly impractical!

7

Implementation: Paradigm 2

• Paradigm 2: Precomputed solutions based on off-line optimization• Find properties of the solution suited for simple and robust on-line

implementation

• Examples– Marathon runner– Hierarchical decomposition– Optimal control – Explicit MPC

8

Example paradigm 2: Marathon runner

c = heart rate

select one measurement

• Simple and robust implementation• Disturbances are indirectly handled by keeping a constant heart rate• May have infrequent adjustment of setpoint (heart rate)

9

Example paradigm 2: Optimal operation of chemical plant

• Hierarchial decomposition based on time scale separation

Self-optimizing control: Acceptable operation (=acceptable loss) achieved using constant set points (cs) for the controlled variables c

cs

• No or infrequent online optimization.• Controlled variables c are found based on off-line analysis.

10

Example paradigm 2: Feedback implementation of optimal control (LQ)

• Optimal solution to infinite time dynamic optimization problem

• Originally formulated as a “open-loop” optimization problem (no feedback)

• “By chance” the optimal u can be generated by simple state feedbacku = KLQ x

• KLQ is obtained off-line by solving Riccatti equations

• Explicit MPC: Extension using different KLQ in each constraint region

• Summary: Two paradigms MPC1. Conventional MPC: On-line optimization

2. Explicit MPC: Off-line calculation of KLQ for each region(must determine regions online)

11

Example paradigm 2: Explicit MPC

MPC: Model predictive controlNote: Many regions because of future constraints

A. Bemporad, M. Morari, V. Dua, E.N. Pistikopoulos, ”The Explicit Linear Quadratic Regulator for Constrained Systems”,

Automatica, vol. 38, no. 1, pp. 3-20 (2002).

12

Issues Paradigm 2: Precomputed on-line solutions based on off-line optimization

Issues (expected research results for specific application):

1. Find structure of optimal solution for specific problems• Typically, identify regions where different set of constraints are active

2. Find optimal values (or trajectories) for unconstrained variables

3. Find analytical or precomputed solutions suitable for on-line implementation

4. Find good “self-optimizing” variables c to control in each region• Find good variable combinations to control

5. Determine a switching policy between different regions

13

• The ideal “self-optimizing” variable is the gradient (first-order optimality condition (ref: Bonvin and coworkers)):

• Optimal setpoint = 0

• BUT: Gradient can not be measured in practice• Possible approach: Estimate gradient Ju based on measurements y

• Here alternative approach: Find optimal linear measurement combination

which when kept constant ( § n) minimize the effect of d on loss. Loss = J(u,d) – J(uopt,d); where u is input for c = constant § n

• Candidate measurements (y): Include also inputs u

How find “self-optimizing” variable combinations in a systematic manner?

Unconstrained degrees of freedom:

14

B. Optimal measurement combination

H

Unconstrained degrees of freedom:

15

Amazingly simple!

Sigurd is told how easy it is to find H

B. Optimal measurement combination

B1. Nullspace method for n = 0 (Alstad and Skogestad, 2007)

Basis: Want optimal value of c to be independent of disturbances

• Find optimal solution as a function of d: uopt(d), yopt(d)

• Linearize this relationship: yopt = F d

• Want:

• To achieve this for all values of d:

• To find a F that satisfies HF=0 we must require

• Optimal when we disregard implementation error (n)

V. Alstad and S. Skogestad, ``Null Space Method for Selecting Optimal Measurement Combinations as Controlled Variables'', Ind.Eng.Chem.Res, 46 (3), 846-853 (2007).

Unconstrained degrees of freedom:

17

B. Optimal measurement combination

B2. Combined disturbances and implementation errors (“exact local method”) Loss L = J(u,d) – Jopt(d). Keep c = Hy constant , where y = Gyu + Gy

dd + ny

• I.J. Halvorsen, S. Skogestad, J.C. Morud and V. Alstad, ``Optimal selection of controlled variables'', Ind. Eng. Chem. Res., 42 (14), 3273-3284 (2003).

Optimization problem for optimal combination:

Theorem 1. Worst-case loss for given H (Halvorsen et al, 2003):

Unconstrained degrees of freedom:

Applies to any H (selection/combination)

18

B. Optimal measurement combination

V. Alstad, S. Skogestad and E.S. Hori, ``Optimal measurement combinations as controlled variables'', Journal of Process Control, 18, in press (2008). V. Kariwala, Y. Cao, S. jarardhanan, “Local self-optimizing control with average loss minimization”, Ind.Eng.Chem.Res., in press (2008)

F – optimal sensitivity matrix = dyopt/dd

B2. Exact local method for combined disturbances and implementation errors.

Theorem 2. Explicit formula for optimal H. (Alstad et al, 2008):

Theorem 3. (Kariwala et al, 2008).

Unconstrained degrees of freedom:

19

Toy Example

20

Toy Example: Single measurements

Want loss < 0.1: Consider variable combinations

Constant input, c = y4 = u

21

Toy Example: Measurement combinations

22

B1. Nullspace method (no noise)

• Loss caused by measurement error only• Recall rank single measurements: 3 > 2 > 4 > 1

23

B2. Exact local method (with noise)

24

B2. Exact local method, 2 measurements

Combined loss for disturbances and measurement errors

25

B2. Exact local method, all 4 measurements

26

Example: CO2 refrigeration cycle

Unconstrained DOF (u)Control what? c=?

pH

27

CO2 refrigeration cycle

Step 1. One (remaining) degree of freedom (u=z)

Step 2. Objective function. J = Ws (compressor work)

Step 3. Optimize operation for disturbances (d1=TC, d2=TH, d3=UA)• Optimum always unconstrained

Step 4. Implementation of optimal operation• No good single measurements (all give large losses):

– ph, Th, z, …

• Nullspace method: Need to combine nu+nd=1+3=4 measurements to have zero disturbance loss

• Simpler: Try combining two measurements. Exact local method:

– c = h1 ph + h2 Th = ph + k Th; k = -8.53 bar/K

• Nonlinear evaluation of loss: OK!

28

Refrigeration cycle: Proposed control structure

Control c= “temperature-corrected high pressure”

29

Summary: Procedure selection controlled variables

1. Define economics and operational constraints

2. Identify degrees of freedom and important disturbances

3. Optimize for various disturbances

4. Identify active constraints regions (off-line calculations)

For each active constraint region do step 5-6:

5. Identify “self-optimizing” controlled variables for remaining degrees of freedom

6. Identify switching policies between regions

30

Example switching policies – 10 km

1. ”Startup”: Given speed or follow ”hare”

2. When heart beat > max or pain > max: Switch to slower speed

3. When close to finish: Switch to max. power

Another example: Regions for LNG plant(see Chapter 7 in thesis by J.B.Jensen, 2008)

31

Current research 1 (Sridharakumar Narasimhan and Henrik Manum):

Conditions for switching between regions of active constraints

Idea:• Within each region it is optimal to

1. Control active constraints at ca = c,a, constraint

2. Control self-optimizing variables at cso = c,so, optimal

• Define in each region i:

• Keep track of ci (active constraints and “self-optimizing” variables) in all regions i

• Switch to region i when element in ci changes sign• Research issue: can we get lost?

32

Current research 2 (Håkon Dahl-Olsen):

Extension to dynamic systems

• Basis. From dynamic optimization:– Hamiltonian should be minimized along the trajectory

• Generalize steady-state local methods:– Generalize maximum gain rule

– Generalize nullspace method (n=0)

– Generalize “exact local method”

33

Current research 3 (Sridharakumar Narasimhan and Johannes Jäschke):

Extension of noise-free case (nullspace method) to nonlinear systems

• Idea: The ideal self-optimizing variable is the gradient

• Optimal setpoint = 0

• Certain problems (e.g. polynomial)– Find analytic expression for Ju in terms of u and d

– Derive Ju as a function of measurements y (eliminate disturbances d)

34

• Our results on optimal measurement combination (keep c = Hy constant)• Nullspace method for n=0 (Alstad and Skogestad, 2007)

• Explicit expression (“exact local method”) for n≠0 (Alstad et al., 2008)

• Observation 1: Both result are exact for quadratic optimization problems

• Observation 2: MPC can be written as a quadratic optimization problem and optimal solution is to keep c = u – Kx constant.

• Must be some link!

Current research 4 (Henrik Manum and Sridharakumar Narasimhan):

Self-optimizing control and Explicit MPC

35

Quadratic optimization problems• Noise-free case (n=0)

• Reformulation of nullspace method of Alstad and Skogestad (2007) – Can add linear constraints (c=Hy) to quadratic problem with no loss

– Need ny ≥ nu + nd. H is unique if ny = nu + nd (nym = nd)

– H may be computed from nullspace method,

•V. Alstad and S. Skogestad, ``Null Space Method for Selecting Optimal Measurement Combinations as Controlled Variables'', Ind. Eng. Chem. Res, 46 (3), 846-853 (2007).

36

Quadratic optimization problems• With noise / implementation error (n ≠ 0)

• Reformulation of exact local method of Alstad et al. (2008) – Can add linear constraints (c=Hy) with minimum loss.

– Have explicit expression for H from “exact local method”

•V. Alstad, S. Skogestad and E.S. Hori, ``Optimal measurement combinations as controlled variables'', Journal of Process Control, 18, in press (2008).

37

Optimal control / Explicit MPC• Treat initial state x0 as disturbance d. Discrete time constrained MPC

problem:

• In each active constraint region this becomes an unconstrained quadratic optimization problem ) Can use above results to find linear constraints

1. State feedback with no noise (LQ problem) Measurements: y = [u x]

Linear constraints: c = H y = u – K x

• nx = nd: No loss (solution unchanged) by keeping c = 0, so u = Kx optimal!

• Can find optimal feedback K from “nullspace method”, – Same result as solving Riccatti equations

• NEW INSIGHT EXPLICIT MPC: Use change in sign of c for neighboring regions to decide when to switch regions •H. Manum, S. Narasimhan and S. Skogestad, ``A new approach to explicit MPC using self-optimizing control”, ACC, Seattle, June 2008.

38

Explicit MPC. State feedback.Second-order system

time [s]

Phase plane trajectory

39

Optimal control / Explicit MPC

2. Output feedback (All states not measured). No noise

• Option 1: State estimator

• Option 2: Direct use of measurements for feedback

“Measurements”: y = [u ym]

Linear constraints: c = H y = u – K ym

• No loss (solution unchanged) by keeping c = 0 (constant), so u = Kym is optimal, provided we have enough independent measurements:

ny ≥ nu + nd ) nym ≥ nd

• Can find optimal feedback K from “self-optimizing nullspace method”

• Can also add previous measurements, but get some loss due to causality (cannot affect past outputs)

H. Manum, S. Narasimhan and S. Skogestad, ``Explicit MPC with output feedback using self-optimizing control”, IFAC World Congress, Seoul, July 2008.

40

Explicit MPC. Output feedbackSecond-order system

time [s]

State feedback

41

Optimal control / Explicit MPC

2. Output feedback – Further extensions

• Explicit expressions for certain fix-order optimal controllers

• Example: Can find optimal multivariable PID controller by using as “measurements”

• Current output (P)

• Sum of outputs (I)

• Change in output (D)

42

Optimal control / Explicit MPC

3. Further extension: Output feedback with noise

• Option 1: State estimator

• Option 2: Direct use of measurements for feedback

“Measurements”: y = [u ym]

Linear constraints: c = H y = u – K ym

• Loss by using this feedback law (adding these constraints) is minimized by computing feedback K using “exact local method”

H. Manum, S. Narasimhan and S. Skogestad, ``Explicit MPC with output feedback using self-optimizing control”, IFAC World Congress, Seoul, July 2008.

43

Conclusion

• Simple control policies are always preferred in practice (if they exist and can be found)

• Paradigm 2: Use off-line optimization and analysis to find simple near-optimal control policies suitable for on-line implementation

• Current research: Several interesting extensions– Optimal region switching

– Dynamic optimization

– Explicit MPC

• Acknowledgements

– Sridharakumar Narasimhan

– Henrik Manum

– Håkon Dahl-Olsen

– Vinay Kariwala