paris’09 daniel r. ramirez & eduardo f. camacho min ... eduardo camacho/wednesday...4...
TRANSCRIPT
1
1Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-Max Model PredictiveControl
Implementation StrategiesDaniel R. Ramirez & Eduardo F. Camacho
2Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Outline
Why Min-Max Model Predictive Control ?
Min-Max Model Predictive Control using open-
loop predictions.
Min-Max Model Predictive Control using closed-
loop predictions.
3Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Why Min-Max Model Predictive Control ?
Model Predictive control is a successboth the industrial community andacademia:
4600 Applications (Qin y Badgwell, 2002).
67 % Refining and petrochemical industry.
Hundreds of paper published.
4Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Why Min-Max Model Predictive Control ?
• Pros:Multivariable from the beginning.Deadtime compensation.Optimal design.Guaranteed stability (academic version).Able to consider constraints.
• Cons:Computational burden.Need of a model.
Modelling errors (uncertainties, disturbances) can degrade the closed loop performance.Modelling errors (uncertainties, disturbances) can degrade the closed loop performance.
RobustModelPredictiveControl
5Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Why Min-Max MPC ?
• Robust MPC:– Robust stability guaranteed:
• Keep the state within a region.• Robust constraint satisfaction.
– Improve performance in presence of uncertaintiesand disturbances:
• Improving robustness.• Take into account uncertainties in the design.
Min-Max formulation: worstcase optimized designMin-Max formulation: worstcase optimized design
0 1 2 3 4 5 6-0.2
0
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
Min-Max MPC (MMMPC)Campo y Morari, 1987
6Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Why Min-Max Model Predictive Control ?
0 2 4 6 8 10-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1MPC
segundos
y
0 2 4 6 8 10-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1MMMPC
segundos
y
Better performance againstuncertainties More RobustnessBetter performance againstuncertainties More Robustness
2
7Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Why Min-Max Model Predictive Control ?
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
muestras
y k
MMMPC
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
muestras
y k
MPC
This increase in robustness alsoappear in real world applicationsThis increase in robustness alsoappear in real world applications
Control Engineering Practice, 2005
8Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Why Min-Max Model Predictive Control ?
10 20 30 40 50 60 70 80 90-1.5
-1
-0.5
0
0.5
1
muestras
Δ u k
10 20 30 40 50 60 70 80 900
0.2
0.4
0.6
0.8
1
1.2
muestras
y kMore robustness can also helpavoding feasibility problems:
Constraint violation due tomodelling errors.
Unfeasible optimization problem.
More robustness can also helpavoding feasibility problems:
Constraint violation due tomodelling errors.
Unfeasible optimization problem.
9Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Why Min-Max Model Predictive Control ?
4 5 6 710-1
100
101
102
103
horizonte de predicción
Tiem
po (s
eg.)
Even when there are advantagesthe number of reported applicationsis very low
Even when there are advantagesthe number of reported applicationsis very low
¿?
ComputationalBurden issues
Berenguel et al. 1997Kim et al. 1998Álvarez et al. 2003
10Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Why Min-Max Model Predictive Control ?
• Desiderable properties of an Min-Max MPC:– Far lower computational burden, complexity growing
with prediction horizon at non exponential rate.– Use efficient numerical algorithms (such as in QP) to
compute control signal.– Flexible implementation, able to change parameters in
real-time.– Able to use constraints.– Low error if an approximation of the optimizer is used. – Robust stability guaranteed.
11Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Types of Min-Max MPC
• A broad division of Min-Max MPC can be established between thosewhich use open-loop predictions and those with closed-looppredictions (note that despite its type all of them are feedback controllers).
• Open-loop prediction based controllers are the most maturealgorithms dating back to 1987 (Campo and Morari).
• They suffer from conservatism in the predicted evolution of the processthat can lead to poor performance or feasibility problems. However, properly tuned they work well as proved later.
• Although its computational burden grows exponentially with certainparameters, there exists techniques to implement it in a wide class ofsystems.
12Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Types of Min-Max MPC
• On the other hand, Closed-loop prediction based controllers havelower conservatism and better feasibility properties, because they takeinto account the fact that the control law is applied in a feedback manner.
• The price to be paid is an even greater computational burden and fewerimplementation options (in fact there is no any single reportedapplication to a real process).
• These controllers date from 1997 (Lee and Yu) and 1998 (Scokaert et al) and usually rely in recursive min-max problems or the optimizationof a set control policies instead of a sequence of control signal values.
• Finally there is a comprimise between these two types of controllerscalled the semi-feedback approach, that can be used with open-loopprediction controllers to add a certain degree of closed-loop behavior in the predictions.
3
13Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Open loop vs close loop prediction
14Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Open loop vs close loop prediction
15Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Open loop vs close loop prediction
Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-Max Model Predictive Control basedon Open-Loop Predictions
17Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Open-Loop Predictions
Robust Model Predictive ControlExplicit implementation
Implementation using a nonlinear bound
Implementation using a QP based bound
Applications and Examples
18Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Predictive Control
Output predictions
Manipulated
y(t+k)
variables
t t+1 t+N
futurepast
u(t+k)
Apply the first component of u*(t)
Solve a finite time optimal control problem to obtain u*(t)
Get new measures y(t)
Estimate x(t)
At each sampling time t:
4
19Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Robust Predictive Control
Output Predictions
VariablesManipulated
t t+1 t+N
futurepast
u(t+k)
y(t+k)
Based on models that take into account uncertainties:
UNCERTAIN PREDICTIONS
20Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Robust Model Predictive ControlSystem model:
Bounded additive uncertainties
Two strategies to consider u(t):
Open-loop predictions:
Semi-feedback predictions:Computed by the controller
21Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Robust MPC
Outputpredictions
VariablesManipulated
t t+1 t+N
futurepast
u(t+k)
y(t+k)
min-maxMPC Plant C
K
w v u xy
-+
The inner loop pre-stabilizesthe nominal system
22Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (∞-∞ norm)
• Campo and showed that by using an ∞-∞norm themin-max problem reduces to a linear programmingproblem.
• Although the algorithm was developed for thetruncated impulse response, it can easily be extended to the other models used.
• The objective function is now described as
23Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (∞-∞ norm)
24Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (∞-∞ norm)
5
25Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (∞-∞ norm)
26Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (∞-∞ norm)
A LP (with many constraints: the vertices of the uncertainty polytope)
27Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (1-norm)
28Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (1-norm)
29Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (quadratic cost)
Cost Function:
Constraints:
30Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
The Min-Max Strategy:
Robust constraintSatisfaction
V*(x,u) is convex
Dependence on uncertainty in constraints can be removed offline
(Lofberg 03,Kerrigan 03)The “max” function:
Convex function
The maximum is attained at least at one of the vertices of Θ.(Bazaraa and Shetty 79)
Min-max MPC (quadratic cost)
6
31Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max MPC (quadratic cost)
Thus, the max function can be computed as:
Number of vertices: Exponential complexity
NP hard problem
As a result, the min-max problem is of the NP hard kind and in general it cannot be solved in real time except for slow processes and/or small horizons.
32Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Real time implementation of Min-Max MPC
Use an explicit implementation, in which the min-
max problem is precomputed for feasible states.
Solve an equivalent reduced min-max problem.
Substitute the worst case cost for a low
computational burden close approximation.
How to get an Min-Max MPC that can be computed in real time ??
Some suggestions:
33Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Open-Loop Predictions
Robust Model Predictive Control
Explicit implementationImplementation using a nonlinear bound
Implementation using a QP based bound
Applications and Examples
34Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Explicit implementation of Min-Max MPC
MMMPC not solvable on line
An alternative: Multi parametric programming
Explicit form control laws:
35Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Multi-parametric Programming
Successfully applied to MPC:
YESYESYESLP based norms
NOYesYESQuadratic norms
Min Max Closed Loop predictions
Min MaxOpen Loop predictions
Nominal
(Bemporad et.al. AUT 02, Bemporad et.al. TAC 03)
(Sakizlis et.al. AUT 04, Wan et.al. TAC 04)
Approximate closed loop explicit solutions
36Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Geometrical Approach
(Ramirez and Camacho Automatica 2006); (Muñoz de la Peña et.al. CEP 05, SCL 2006)
• Based on properties of cost function and the concept of active vertices (i.e., the vertices in which the worst case is attained at the solution).
• It allows:
• To prove that the control law is PWA on the states.
• To find an explicit solution using a constructive algorithm.
Works well, but… difficult to understand, complex specially with constraints, heuristic if need to be efficient.
7
37Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max as a Quadratic Program
The min-max problem can be rewritten as:
Added and subtracted V(x,v,0)
Taking into account the cost definition:
Quadratic dependence on wAffine dependence on x and v
38Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max as a Quadratic Program
Min max problem:
Epigraph approach:
Maximization
39Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-max as a Quadratic Program
Min max problem:
Quadratic Programming problemExponential number of constraintsSemi-definite positive function
MULTIPARAMETRIC THEORY
(Tondel et.al. CDC 03)
linear constraints
40Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Example
Consider the double integrator with bounded additive uncertainties:
The uncertainty satisfies:
The state and control input are constrained:
The control performance objectives are described by
A linear feedback law is considered:
41Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Example
−10 −8 −6 −4 −2 0 2 4 6 8 10−10
−8
−6
−4
−2
0
2
4
6
8
10
x1
x 2
N = 1
I=1 A=10
I=1 A=[ ]
I=2 A=[ ]
I=1 A=9
I=2 A=10
I=2 A=9
−10 −8 −6 −4 −2 0 2 4 6 8 10−10
−8
−6
−4
−2
0
2
4
6
8
10
x1
x 2
N = 1
I=1 A=10
I=1 A=[ ]
I=2 A=[ ]
I=1 A=9
I=2 A=10
I=2 A=9
N=1
42Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Example
−10 −8 −6 −4 −2 0 2 4 6 8 10−10
−8
−6
−4
−2
0
2
4
6
8
10
x1
x 2
N=3
−10 −8 −6 −4 −2 0 2 4 6 8 10−10
−8
−6
−4
−2
0
2
4
6
8
10
x1
x 2
N=3
−10 −8 −6 −4 −2 0 2 4 6 8 10−10
−8
−6
−4
−2
0
2
4
6
8
10
x1
x 2
N=5
−10 −8 −6 −4 −2 0 2 4 6 8 10−10
−8
−6
−4
−2
0
2
4
6
8
10
x1
x 2
N=5
N=3
N=5
8
43Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Example
−10 −8 −6 −4 −2 0 2 4 6 8 10−10
−8
−6
−4
−2
0
2
4
6
8
10
x1
x 2
N = 1
I=1 A=10
I=1 A=[ ]
I=2 A=[ ]
I=1 A=9
I=2 A=10
I=2 A=9
−10 −8 −6 −4 −2 0 2 4 6 8 10−10
−8
−6
−4
−2
0
2
4
6
8
10
x1
x 2
N = 1
I=1 A=10
I=1 A=[ ]
I=2 A=[ ]
I=1 A=9
I=2 A=10
I=2 A=9
Set of active verticesis very small
44Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Feedback PT-326
Heat Stored
Heat In
Heat Out
•Scaled laboratory process•2nd order system•Fast dynamics•Ts = 0.4 s.
(Explicit solution)
45Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Results of the MMMPC
0 100 200 300 400 5002
4
6
8
0 100 200 300 400 5000
5
10
0 100 200 300 400 5002
4
6
8
0 100 200 300 400 5000
5
10
0 100 200 300 400 5002
4
6
8
0 100 200 300 400 5000
5
10
0 100 200 300 400 5002
4
6
8
samples0 100 200 300 400 500
0
5
10
samples
MMMPC with linear feedback yk u
k
N=15, λ =1
N=20, λ =15
N=8, λ =10
N=8, λ =5
Different controllers
Applied with success
(Muñoz de la Peña et.al CEP 05)
46Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Comparisons
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
samples
y k
MPC
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
samples
y k
MMMPC
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
samples
y k
MMMPC with linear feedback
MPC
MMMPC
MMMPCwith linear feedback
Different positions of the inlet throttle from 20o to 100o
47Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Explicit implementation: Pros and Cons
• Possibility to implement explicit min-max MPC control laws. using standard multi-parametric techniques.• Well suited for fast systems and embedded applications.• But…⌧ The exponential number of constraints limits the value of the prediction horizon.⌧ The implementation is not flexible, the user cannot change parameters without recomputing the controller.
48Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Open-Loop Predictions
Robust Model Predictive Control
Explicit implementation
Implementation using a nonlinear boundImplementation using a QP based bound
Applications and Examples
9
49Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Using an upper bound of the worst case cost
A computationally cheaper upper bound is minimized instead of the worst case cost :
Previous results:LMI methods Kothare et.al 96, Lofberg 03BQP Algorithms Alamo el.al 04
Lowercomputationalburden
Strategy based in a nonlinear bound, Ramirez et al., JPC 2006
Strategy based in QP problems, Alamo et al., Automatica 2007
Strategy based in a nonlinear bound, Ramirez et al., JPC 2006
Strategy based in QP problems, Alamo et al., Automatica 2007
50Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Strategy based on a nonlinear bound
Main Properties:Low computational burden.Based on matrix computations.Close to the worst case cost.
Main contribution:
New upper bound of the maximization problem not based on LMIs or complex algorithms.
51Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Computing the nonlinear upper bound
The quadratic maximization problem can be rewritten as:
Asume that ε = 1, then:Binary optimization problem
Note that matrix H depends on (x,u)
52Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Computing the nonlinear upper bound
The idea to obtain the bound is:
If T is a diagonal positive definite matrix such that:
then:
Computation of T: LMI orProposed method based on simple matrix computations
53Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Computing the nonlinear upper bound
T will be obtained adding to H a series of matrices of the form:
The vi will be computed in such a way that allow to obtain a close bound with low complexity.
positive definite matrix
54Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Computing the nonlinear upper bound
Vectors vi are computed so that matrix H is partially diagonalized:
If
This idea is applied recursively
Free parameter
10
55Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
cc
Computing the nonlinear upper bound
The error introduced at each diagonalization step must be minimum:
56Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Nonlinear upper bound algorithm
Let
For k=1 to n-1
Let Hsub = [Tij] for i,j= k,...,n Compute α for Hsub
Form
Form
End For
When the procedure is over, the bound is computed as σ u(H) = trace(T)
57Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Computing the control signal
• At each sampling time, solve:
and apply the first component of u*(x) using a receding horizon strategy.
If at any step of the diagonalization procedure, all the elements of Hsubare ≥ 0 the best bound will be
No need to continuethe computationloop
58Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Accuracy of the nonlinear upper bound
-1-0.5
00.5
1
20
30
40
500
10
20
30
40
media de los elementos de H0
Dimension matriz
Des
viac
ión
de la
cot
a LM
I (%
)
It will be compared against theLMI bound:
Nesterov et al, 2000
Deviation for a set of random matrices of the form H=H0'H0.
Typical matrices in MPC problems are outside this zone, because the mean of its elements are different from zero
59Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Computational burden of the bound
Relative speed up from the LMI bound Solver LMI by F. Rendl
http://www.math.uni-klu.ac.at/or/Software.
-2-1
01
2
20
30
40
500
200
400
600
800
media de loselementos de H0
Dimensión matriz
Ace
lera
ción
Typical matrices in MPC problems are outside this zone, because the mean of its elements are different from zero
60Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Open-Loop Predictions
Robust Model Predictive Control
Explicit implementation
Implementation using a nonlinear bound
Implementation using a QP based boundApplications and Examples
11
61Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
QP based strategy
• It has two desiderable properties:√ Robust stability guaranteed.√ Implementation based on an optimization problem similar to that
of conventional MPC Quadratic Programming (QP).• Outline of the strategy:
1. Obtain an initial estimation of the solution of the min-max problem.
2. Obtain a quadratic function that bounds the worst case cost.
2.1 Compute some parameters denoted αk for the initial estimation.
2.2 Use a diagonalization procedure to obtain the quadratic function.
3. Compute the control signal as the minimizer of the quadratic function.
62Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
QP based strategy
Robust stability guaranteed, but requires semi-feedback for non open-loop stable sytems:
The control law of the original MMMPC original is computed by solving at each step t:
Stability conditions imposedon Terminal Cost and Region
Terminal region
IncludesIncludes a a terminal terminal costcost
StabilizingdesignMayne, 2000
63Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Stabilizing design
If there exists a solution for x(k)
Then existssolution for x(k+1)
Convergence to the origin(Optimal cost Lyapunovfunction)
Conditions on the terminal regionAdmissiblerobustinvariant set
Maximal robust invariant set computed for the systema controlled by:
Condition on the terminal cost
Guaranteesrobustfeasibility
64Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
QP problems based strategy
Step 1: Initial estimation of the solution of the min-max problem
The maximum cost can be computed as:
An easily computable (although not very good) upper bound is :
Initialestimation
QP equivalentSlack variables
65Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Step 2: Computing the quadratic function
Step 2.1: Compute the parameters αk for the initial estimationTaking into account that if it is assumed ε=1:
Use the diagonalization algorithm of the nonlinear bound overto obtain the α k:
The resulting diagonal matrix, denoted by σu(v) verifies:
66Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Step 2: Computing the quadratic function
• Step 2.2: Compute a matrix denoted by
12
67Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Step 3: Computing the signal control• Let:
• The control signal will be computed solving:
• The suboptimality is bounded:
68Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Stability• The proposed control law guarantees the closed loop robust stability. • Outline of the proof:
– Let J(x(t)) be the optimal cost of the original min-max problem.– It can be proved that using the proposed control law there exists γ > 0
such that:
– This implies that:
– That is, from any x(t) the state evolves into:
It can escape
69Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Stability• The state is always steered into Φε, but it can be escape from it.• The set to which the state can evolve from Φε is:
70Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Open-Loop Predictions
Robust Model Predictive Control
Explicit implementation
Implementation using a nonlinear bound
Implementation using a QP based bound
Applications and Examples
71Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
A simulated example
F1
F2
FR
h1
h2
o1=a1h1
o2=a2h2
Two tank process:Ogunnaike y Ray, 1994
Tanks sections: 3 m2 and 2m2.a1 = a2 = 0.5 m2min-1. 40% of the liquid is pumped back
to tank 1.
72Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
A simulated example
Controller based on the nonlinear bound
ε = 0.02N=15, Nu=10, 230 vertices
0 20 40 60 80 100 1200
0.2
0.4
nive
les
0 20 40 60 80 100 1200
0.2
0.4
0.6
0.8
muestras
caud
ales
13
73Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Computational burden
Average, max and min speed-up over the exact MMMPC.
4 6 8 10 12 14 16 18 200
100
200
300
400
500
600
700
800
900
horizonte de predicción
Ace
lera
ción
(med
ia y
máx
-mín
)
4 5 6 7 8 910
0
101
102
103
104
105
106
horizonte de predicción
Ace
lera
ción
(med
ia y
máx
-mín
)
Average, max and min speed-up over the minimization of the LMI bound.
74Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Accuracy of the nonlinear bound
Desviation from the exact optimal cost and from the optimal cost obtained using the LMI bound.
75Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Accuracy of the nonlinear bound
1 20 40 60 80 100 1200
1
2
3
4
5
6
7
muestras
cost
e óp
timo
1 20 40 60 80 100 1200
5
10
15
muestras
cost
e óp
timo
Optimal cost with the nonlinear bound and the LMI bound for a simulation (Nu=5,N=20).
Optimal cost with the nonlinear bound and the exact MMMPC for a simulation.
76Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
A simulated example
Controller based on QP problems
ε = 0.025N=7, Nu=7
0 20 40 60 80 1000
0.7
1
nive
les
0 20 40 60 80 1000
0.2
0.4ca
udal
es
10 20 30 40 50 60 70 80 90 100-0.03
-0.02
-0.01
0
muestras
desv
iaci
ón
77Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
A simulated example
10 20 30 40 50 60 70 80 90 1000
2
4
6
8
10
12
14
muestras
cost
es ó
ptim
os
10 20 30 40 50 60 70 80 90 1000
5
10
cost
es ó
ptim
os
10 20 30 40 50 60 70 80 90 1000
1
2
3
4
muestras
desv
iaci
ón (%
)
Optimal cost of the exactMMMPC and the QP basedcontroller (Nu=N=7).
Optimal cost of the exact MMMPC when it is used both the exact solutionand that of the QP based controller, that is
78Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
A simulated example
4 5 6 7 8 9100
101
102
103
104
105
106
horizonte de predicción
Ace
lera
ción
(m
edia
y m
áx-m
ín)
Average, max and min speed-up for the QP basedcontroller over the exact MMMPC.
14
79Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Application to a pilot plant
Inte
rcam
biad
or
de C
alor
Tank with a 14.4kW heater.
Recirculation through a heat exchanger.
Industrial instrumentation.
Connected to Simatic-IT.
It emulates a CSTR using the heater tosimulate the heat of an exothermic reaction.
Model by Santos et al., 2001
80Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Application to a pilot plant
0 100 200 300 400 500 600
40
60
80
TT5
o C
0 100 200 300 400 500 600
40
60
80
Tiempo (minutos)
V8
%
0 100 200 300 400 500 600-0.5
-0.25
0
0.25
0.5
Tiempo (minutos)
erro
r
Multilevel Step sequence
Identification of a first ordermodel by Least Squares
81Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Application to a pilot plant
0 25 50 75 100 125 15035
40
45
50
55
60
65
TT5
o C
0 25 50 75 100 125 15020
40
60
80
100
Tiempo (minutos)
V8
%
0 25 50 75 100 125 1500.1
0.2
0.3
Tiempo (minutos)
CA
Controller based on the nonlinear boundε = 0.25N=15, Nu=12
Sampling time: 60 seg.
Deadtime rounded up to 1.
Smith Predictor-like correction.Normey y Camacho, 07
82Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Application to a pilot plant
0 20 40 60 80 100 12045
50
55
60
65
TT5
o C
0 20 40 60 80 100 12020
40
60
80
V8
%
0 20 40 60 80 100 1203040506070
u
0 20 40 60 80 100 1200.050.1
0.150.20.2
Tiempo (minutos)
CA
0 20 40 60 80 10040
45
50
55
60
65
TT5
o C
0 20 40 60 80 10030
40
50
60
70
80
V8
%
0 20 40 60 80 1000.05
0.1
0.15
0.20.25
Tiempo (minutos)
CA
Disturbances in the input (15%)Disturbances in the inlet flow (25%)
83Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Application to a pilot plant
Controller based on QP problems
ε = 0.25N=15, Nu=12
0 25 50 75 100 125 15035
40
45
50
55
60
65
T
0 25 50 75 100 125 15020
40
60
80
100
tiempo
V8
%
0 50 100 1500
0.1
0.2
0.3
Tiempo (minutos)
CA
84Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Application to a pilot plant
0 20 40 60 80 100 12040
45
50
55
60
65
T
0 20 40 60 80 100 12030
40
50
60
70
80
Tiempo (minutos)
v8
0 20 40 60 80 100 1200.05
0.1
0.15
0.2
0.25
Tiempo (minutos)
CA
0 25 50 75 100 125 15035
40
45
50
55
60
65
T
0 25 50 75 100 125 15020
40
60
80
100
tiempo
v8
MMMPC NLMPCMMMPC QPref
MMMPC NLMPCMMMPC QP
Disturbances in the inlet flow (25%)
Comparison for a set point trackingexperiment using:
MPC.MMMPC based on the nonlinear
bound.MMMPC based on QP problems.
15
85Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Feedback PT-326
Heat Stored
Heat In
Heat Out
•Scaled laborat ory process•2nd order syst em•Fast dynamics•T s = 0.4 s.
( Explicit solut ion)
86Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
0 500 1000 15000
1
2
3
4
5
6
7
8
9
10
0 500 1000 15000
0.5
1
1.5
2
2.5
3
3.5
0 500 1000 1500−0.4
−0.3
−0.2
−0.1
0
0.1
0.2
0.3
e k
0 500 1000 1500−0.15
−0.1
−0.05
0
0.05
0.1
0.15
0.2
samples
e k
One step identification error for ARX model
One step identification error for CARIMA model
System Identification
A(z−1)yk = B(z−1)uk−1−dARX
L east squares ident if icat ion met hod
Random binary input
Ident if icat ion set / Validat ion set
Error bounding CARIM A model
ΔA(z−1)yk = z−dB(z−1)Δuk−1−d+ θkCA RIM A
B est f it t ing
T s = 0.4D elay = 4
87Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
System Identification
A(z−1) = 1 −0.5510z−1− 0.4072z−2
B(z−1) = 0.0090+ 0.0127z−1+ 0.0105z−2
B est m odel:
D elay = 4error bound = 0.05
A(z−1)yk = B(z−1)uk−1−dARX ΔA(z−1)yk = z−dB(z−1)Δuk−1−d+ θkCA RIM A
88Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Delay Compensation
Smit h Predict or St rat egy:
yk+d= yk+d|k+ (yk|k−d− yk)
PT -326PT -326
G( z)G( z)
z-dz-d+ -
+
+
ukyk
yk+ d,k
yk,k-d
yk+d
89Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Reference Traking
In order t o evalut at e reference t raking, t he st at evect or is augm ent ed:
( B em porad et .al. 2002)
F ixed reference over t he predicit ion horizon
zk =hyk+d−1 yk+d−2 Δuk−1 Δuk−2 Δuk−3 rk
iT
Q uadrat ic crit erion(yk − rk)2 = zTk Qzk
90Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Explicit control law
IF L zk �h4.265 0.232 3.03 3.28 3.28 3.03 0.232 4.265
iTTHEN Δuk =
h−0.006 −0.003 −0.273 0.135 0.104 0.034
izk
L=
⎡⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎣
1.501 0.727 78.847 −34.088 −28.186 −16.5740.023 0.011 1.156 −0.531 −0.426 −0.1990.020 0.010 1.248 −0.443 −0.400 −0.404−0.057 −0.030 −3.521 1.232 1.144 1.1440.057 0.03 3.521 −1.232 −1.144 −1.144−0.020 −0.010 −1.248 0.443 0.400 0.404−0.023 −0.011 −1.156 0.531 0.426 0.199−1.501 −0.727 −78.847 34.088 28.186 16.574
⎤⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎦
16
91Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Results of the MMMPC
0 100 200 300 400 5002
4
6
8
0 100 200 300 400 5000
5
10
0 100 200 300 400 5000
5
10
0 100 200 300 400 5000
5
10
0 100 200 300 400 5000
5
10
0 100 200 300 400 5000
5
10
0 100 200 300 400 5002
4
6
8
samples0 100 200 300 400 500
0
5
10
samples
MMMPC yk u
k
N=16, λ = 10,α=0.95
N=8, λ = 0.3,α=0.8
N=15, λ = 0.5,α=0.75
N=8, λ = 10,α=0.99
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
y k
0 50 100 150 200 250 300 350 400 450 5000
2
4
6
8
samples
u k
MMMPC N=16, λ = 10, α =0.9
“ W ell” t uned M M M PC
A gresive cont rol law
O pen looppredict ions
D if ferent cont rollers
92Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC with Linear Feedback
2 4 6 8 10 12 14−3
−2
−1
0
1
2
3Worst case evolution of a system
Time step k
Out
put y
k
Nominal
Linear feedback
Δuk = Kzk+ vk
L inear feedback:
L inear cont roller
Feedback on t he predict ions wit hout complexit y increase( B emporad et ,al, 1998)
93Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC with Linear Feedback
• ìΔuk = Kzk+ vk
L inear feedback:
PT -326PT -326Δυk yk
vk
K zK z
M M M PCM M M PC +
Pre-cont rolled syst em ( linear model)
94Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Results of the MMMPC with linear feedback
0 100 200 300 400 5002
4
6
8
0 100 200 300 400 5000
5
10
0 100 200 300 400 5002
4
6
8
0 100 200 300 400 5000
5
10
0 100 200 300 400 5002
4
6
8
0 100 200 300 400 5000
5
10
0 100 200 300 400 5002
4
6
8
samples0 100 200 300 400 500
0
5
10
samples
MMMPC with linear feedback yk u
k
N=15, λ =1
N=20, λ =15
N=8, λ =10
N=8, λ =5
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
y k
0 50 100 150 200 250 300 350 400 450 5003
4
5
6
7
samples
u k
MMMPC with linear feedback, N=8, λ = 20, α =1
“ W ell” t uned M M M PC
“ Soft er” cont rol law
Feedbackin t he
predict ions
D if ferent cont rollers
95Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Comparisons
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
samples
y k
MPC
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
samples
y k
MMMPC
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
samples
y k
MMMPC with linear feedback
M PC
M M M PC
M M M PCwit h linear feedback
D if ferent posit ions of t he inlet t hrot t le f rom 20o t o 100o
96Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Comparisons
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
samples
y k
MMMPC
0 50 100 150 200 250 300 350 400 450 5002
3
4
5
6
7
8
9
10
samples
y k
MMMPC with linear feedback
0 50 100 150 200 250 300 350 400 450 5000
2
4
6
8
10
u k
0 50 100 150 200 250 300 350 400 450 5000
2
4
6
8
10
samples
u k
MMMPC
MMMPC with linear feedback
Cont rol Input
17
Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Min-Max Model Predictive Control usingClosed-Loop Predictions
98Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Closed-Loop Predictions
Feedback Min-Max MPCApproximate Multi-parametric Programming
Decomposition algorithm
Example
99Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Feedback Min Max MPC
subject to:
Open Loop Min-Max MPC
Feedback Min-Max MPC
Notion of FEEDBACK(Receding horizon scheme)
100Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Feedback Min Max MPC
subject to:
Feedback Min-Max MPC
Hard problem
References:
•LMI approaches•Multi-parametric approaches•Large scale deterministic equivalents
(Bemporad et.al. TAC 2003)
(Kothare et.al. Automatica 1996)
(Scockaert et.al. TAC 1999)
101Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
How to implement it ??
• Approximate Multi-parametric Programming • Decomposition Algorithm
Clearly an open problem. Suggestions:
102Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
The Closed-Loop Min-Max MPC
• Dynamic Programming can be used to state theclosed-loop min-max MPC.
• The problem can be expressed as the recursiveproblem.
J*t (x (t)) Jt (x (t), u (t)) s.t.
Rx x (t) + Ruu (t) ≥ rF (x (t), u (t), Ø (t)) є X (t + 1)
Jt (x (t), u (t) L (x (t), u (t)) + J*t+1 (x (t + 1))
• Where X (t+1) is the region where functionJ*t+1 (x (t+1) is defined.
18
103Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
The Closed-Loop Min-Max MPC (2)
• x (t+1) = A (Ø (t) x (t) + B (Ø (t)) u (t) + E (Ø (t)), • with • A (Ø (t)) = Ao + Ai Øi (t), • B (Ø (t)) = B0 + BiØi (t), • E (Ø (t)) = E0 + EiØi (t).
• Where Øi (t) denotes the ith component of the uncertainty vector Ø(t) є Θ.
• The stage cost of the objective function defined as:L (x (t+j), u (t+j) ||Q x (t+j) ||p + ||Ru (t+j) ||p
• With the terminal cost defined asJ*t+N (x (t+N)) ||Px (t+N) ||p.
• The solution is a piece affine function of the state.(Bemporad et. al. 2003)
104Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
The Closed-Loop Min-Max MPC (3)
The demostration is based on the following
1. If J (x (t), u) is a convex piecewise affine function(i.e., J (x (t), u) = maxi=1,.. s {Liu + Hi x (t) + ki}),the problem:
is equivalent to the following mp-LP problem:
105Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
The Closed-Loop Min-Max MPC (4)
2. If J (u, x (t),θ) and g (u, x (t), θ) are convex functions in θ for all (u, x (t)) with θ є Θ, where Θis a polyhedron with vertices θ i (i= 1, …, NØ) then the problem
is equivalent to the problem:
106Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
The Closed-Loop Min-Max MPC (5)
3. If J (u, x (t), θ) is convex and piecewise in x (t) u(i.e. J (u, x (t), θ) = Li (θ) u + Hi (θ) x (t) + ki (θ))
4. and g (u, x (t), θ) is affine in x (t) and u (i.e. g (u, x (t), θ) = Lg (θ) u + Hg (θ) x (t) + kg (θ))
5. with Li, Hi, ki, Lg, Hg, kg convex functions, 6. Then the min-max Problem is equivalent to the
problem:
107Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
The Closed-Loop Min-Max MPC (6)
• Let us now consider the first step of thedynamic programming problem (11.36) with
L (x, u ) || Qx (t +j) ||p + || Ru (t+j) ||p• The terminal cost
• and the linear system
108Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
The Closed-Loop Min-Max MPC (7)
Appying the previous results, this is equivalent to the following mp-LP problem:
19
109Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
• J*N-1 is the solution of the mp-LP problem whichresults in a piecewise affine function of x (N-1).
• The corresponding control signal u* (N-1) is also a continuours piecewise affine function of the state.
• The feasible set X (N-1) is a convex polyhedron.• The same recursively for j= N-2, N-3, … 0,• We can conclude that u* (t) is a piecewise affine
function of state x (t). • Notice that the argument does not hold when the
objective funcion is quadratic as funtion (11.42) would not be piecewise affine convex with respect tothe maximization variables θ (N-1)
The Closed-Loop Min-Max MPC (8)
110Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Fast Implementation of MPC and Dead Time Considerations
• The number of regions grows exponentiallywith the horizon, dimension of state vector and number of vertices of uncertainpolytope.
• The state vector dimension which can be very high for processes with long deadtimes. (as can be found frequently in industry)
• In the case of a dead time of d samplinginstants, an augmented state vector
• xa (t) = [x (t)t u (t-1)T… u(t-d)T].
111Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Fast Implementation of MPC and Dead Time Considerations(2)
• To overcome these problems.– Use the predicted state
to compute instead of
• Consider a process modelled by the reactioncurve method with a dead time equal to itstime constant. If the sampling time chosen isone-tenth of the time constant, then dim (xa(t)) = 11 while dim (x (t))= 1
112Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Closed-Loop Predictions
Feedback Min-Max MPC
Approximate Multi-parametric ProgrammingDecomposition algorithm
Example
113Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Multi-parametric Convex Programming
z=optimization variables, x=parameters
Convex constraints
Convex objective function
Feedback MPC can be often solved by convex optimization:
Approximate multiparametric convex programming solver:
A. Bemporad and C. Filippi, Approximate Multiparametric Convex Programming, CDC 2003
114Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Multi-parametric Convex ProgrammingProperties:
piecewise affine (sub-)optimal solution
constraintsatisfaction
(sub-)optimalitywith guaranteederror bound
tree structuresolution
20
115Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Multi-parametric Convex Programming
M.V.Kothare, V. Balakrishna, M.Morari Robust Constrained Model Predictive Control using Linear Matrix Inequalities, Automatica 33, 1996
Linear uncertain systems subject to constraints on (x,u)
Objective function
Application to:
Linear control lawThe gain matrix F is computed at each time step k
116Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Multi-parametric Convex Programming
Kothare´s controller properties:
The control law robustly regulates the system to the origin while assuring robust constraint satisfaction.
Approximate control law:
The control law ultimately bounds the system in a region that contains the origin while assuring robust constraint satisfaction.
(Kothare et.al. Automatica 1996)
(Muñoz de la Peña et.al.TAC 2005)Very efficient implementation
117Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Closed-Loop Predictions
Feedback Min-Max MPC
Approximate Multi-parametric Programming
Decomposition algorithmExample
118Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Decomposition Algorithm
subject to:
Feedback Min-Max MPC
Linear systems with bounded uncertainties
Additive uncertaintiesPolytopic uncertainties
119Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Decomposition Algorithm
Explicit form control laws:
(Bemporad et.al. TAC 2003)Recursive application of mpLP solvers
Complexity depends on state dimension
Cost functions based on a LP problem:
Cost functions based on a QP problem:
No similar result
120Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Decomposition AlgorithmDue to convexity of the prediction model and the cost function:
Only extreme realizations have to be taken into account(finite dimensional problem)
Large scale linear problem(deterministic)
•Solve on-line large LP
•Multi-parametric approach
(Sckocaert et.al. TAC 1998)
(Kerrigan et.al. IJRNC 2004)
Nodes of the tree grow exponentially with N
Time 0 1 2
w=1
w=-1
Root
w=1
w=1
w=-1
w=-1
w(t)=(-1,1)
Cost functions based on a LP problem:
Cost functions based on a QP problem:
No similar result
1
121Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Decomposition Algorithm
Equivalent problem. Multi-Stage Min Max Linear Problem.
Time 0 1 2
w=1
w=- 1
Root
w=1
w=1
w=- 1
w=- 1
w(t)=(-1,1)
Cost-to-go function Vj*
Each node:
Set of variables zj
Depends on uncertainty
Cost functions based on a LP problem:
zj
122Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Decomposition Algorithm
Main contribution
It allows on-line implementation of feedback min-max MPC.(For a broader family of systems)
Algorithm that solves
(Benders, NM 1962)
Based on the ideas introduced by Benders in 1962
MILP problems, LP problems, SP problems
(Muñoz de la Peña et.al. CDC 2005)
123Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Decomposition Algorithm
zi
k = 1k = 2k = 3
Function is PWA Algorithm convergesto optimum
124Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Decomposition Algorithm
Equivalent problem. Multi-Stage Min Max Quadratic Problem.
Time 0 1 2
w=1
w=- 1
Root
w=1
w=1
w=- 1
w=- 1
w(t)=(-1,1)
Allows implementation offeedback min max MPC withquadratic cost criterion.
Approximate solution
Cost functions based on a QP problem:
zj
125Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Decomposition Algorithm
zi
k = 1k = 2k = 3
Function is PWQ
Approximation error
Bound on theerror is a parameter
126Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
MMMPC using Closed-Loop Predictions
Feedback Min-Max MPC
Approximate Multi-parametric Programming
Decomposition algorithm
Example
1
127Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Numerical Examples
Quadruple tank process
nx = 4
Sampling time = 10s.
Model of a real plant
128Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Numerical Examples
Computation times:
It can be implemented up to N=5
More than 2000 nodes
Only algorithm available in the literature.
Problem: Complexity still grows exponentially with N.
Good performance... N=10
Proposed solution: RECOURSE HORIZON – Still an open problem
129Paris’09 Daniel R. Ramirez & Eduardo F. Camacho Min-max MPC
Finally, some papers
D.R. Ramírez and E.F. Camacho, ON THE PIECEWISE NATURE OF MIN-MAX MODEL PREDICTIVE CONTROL WITH BOUNDED UNCERTAINTIES, IEEE CDC’01.
D.R. Ramírez and E.F. Camacho, PIECEWISE AFFINITY OF MIN-MAX MPC WITH BOUNDED ADDITIVE UNCERTAINTIES AND A QUADRATIC CRITERION, Automatica, 2006.
D. Muñoz de la Peña, T. Álamo, D.R. Ramírez y E.F. Camacho, MIN-MAX MODEL PREDICTIVE CONTROL BASED ON A QUADRATIC PROGRAM, IET Proc. Control Theory and Applications, 2007.
T. Álamo, D.R. Ramírez y E.F. Camacho, EFFICIENT IMPLEMENTATION OF CONSTRAINED MIN-MAX MODEL PREDICTIVE CONTROL WITH BOUNDED UNCERTAINTIES: A VERTEX REJECTION APPROACH, Journal of Process Control, 2005.
D.R. Ramírez, T. Álamo, E.F. Camacho y D. Muñoz de la Peña, MIN-MAX MPC BASED ON A COMPUTATIONALLY EFFICIENT UPPER BOUND OF THE WORST CASE COST, Journal of Process Control, 2006.
T. Álamo, D.R. Ramírez, T. Álamo, D. Muñoz de la Peña y E.F. Camacho, MIN-MAX MPC USING A TRACTABLE QP PROBLEM, Automatica, 2007.
D. Muñoz de la Peña, A. Bemporad, and C. Filippi, ROBUST EXPLICIT MPC BASED ON APPROXIMATE MULTI-PARAMETRIC CONVEX PROGRAMMING, IEEE Transactions on Automatic Control, 2006.
David Muñoz de la Peña Sequedo, Teodoro Alamo Cantarero, A. Bemporad, Eduardo Fernández Camacho,A Decomposition Algorithm for Feedback Min-Max Model Predictive Control, IEEE Transactions on Automatic Control, 2006.