Professional Documents
Culture Documents
Introduction Stability, equilibria, invariant sets Lyapunov function theory Disturbances and robust stability Basic MPC problem and nominal stability
James B. Rawlings
2 Department of Chemical and Biological Engineering University of WisconsinMadison 3
Graduate School in Systems, Optimization, Control and Networks (SOCN) K.U. Leuven Leuven, Belgium August 29September 5, 2013
SOCN 2013
1 / 45
SOCN 2013
2 / 45
Predictive control
Reconcile the past
11 00 00 11 1 0 1 0 1 0
State estimation
Forecast the future
00 11 11 00 00 11
MH Estimate Measurement
MH Estimate Measurement
sensors y
1 0 1 0 0 1 1 0 0 1 1 0 0 1 0 1 1 0 00 0 1 11 0 1 0 1 0 1 0 0 1 0 01 1 0 0 1 01 1 0 1 1 0 1 1 0 1 0
1 0 1 0 1 0 0 1 0 1 0 1 1 0 1 0 0 1 0 1 10 0 1 0 1 0 1 0 1 00 1 1 1 0 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 1 0 1 0 01 1 0 1 00 1 0 1 0 1 0 1 1 0 0 1 0 1 1 0 1 0 1 0 0 1
1 0 0 1
1 0 0 1
1 0 0 0 0 0 0 0 0 0 0 0 0 01 1 01 1 01 1 01 1 01 1 01 1 01 1 01 1 01 1 01 1 01 1 0 1
1 0 0 1 1 0 1 0
1 0 0 1
1 0 0 1 1 00 1 1 0 0 1
sensors y
1 0 1 0 0 1 1 0 0 1 1 0 0 1 0 1 1 0 00 0 1 11 0 1 0 1 0 1 0 0 1 0 01 1 0 0 1 01 1 0 1 1 0 1 1 0 1 0
1 0 1 0 1 0 0 1 0 1 0 1 1 0 1 0 1 0 0 1 10 0 1 0 1 0 1 0 1 00 1 1 1 0 1 0 0 1 1 0 0 1 1 0 0 1 1 0 0 1 1 0 1 0 1 0 01 1 0 1 00 1 0 1 0 1 0 1 1 0 0 1 0 1 1 0 1 0 1 0 0 1
1 0 0 1
1 0 0 1
1 0 0 0 0 0 0 0 0 0 0 0 0 01 1 01 1 01 1 01 1 01 1 01 1 01 1 01 1 01 1 01 1 01 1 0 1
1 0 0 1 1 0 1 0
1 0 0 1
1 0 0 1 1 00 1 1 0 0 1
actuators u t time
actuators u t time
min
u (t ) 0
min
x0 ,w (t ) T
|y g (x , u )|2 f (x , u )|2 R + |x Q dt
x = f (x , u ) x (0) = x0
SOCN 2013
x = f (x , u ) + w y = g (x , u ) + v
3 / 45 SOCN 2013
(given)
MPC short course
y = g (x , u )
System model1
We consider systems of the form x + = f (x , u ) where the state x lies in Rn and the control (input) u lies in Rm ; In this formulation x and u denote, respectively, the current state and control, and x + the successor state. We assume in the sequel that the function f : Rn Rm Rn is continuous. Let (k ; x , u) denote the solution of = f (x , u ) at time k if the initial state is x (0) = x and the control sequence is u = {u (0), u (1), u (2), . . .}; The solution exists and is unique.
Most of this preliminary material is taken from Rawlings and Mayne (2009, Appendix B). Downloadable from www.che.wisc.edu/~jbraw/mpc.
SOCN 2013 MPC short course 5 / 45 1
x+
SOCN 2013
6 / 45
In other situations, for example when studying the stability properties of an oscillator, convergence to a specied closed set A Rn is sought. If convergence to a set A is sought, it is desirable for the set A to be positive invariant:
SOCN 2013
7 / 45
SOCN 2013
8 / 45
Denition 4
Dene distance from point x to set A |x |A := inf |x z |
z A
A function : R0 R0 belongs to class K if it is continuous, zero at zero, and strictly increasing; : R0 R0 belongs to class K if it is a class K and unbounded ( (s ) as s ). A function : R0 I0 R0 belongs to class KL if it is continuous and if, for each t 0, (, t ) is a class K function and for each s 0, (s , ) is nonincreasing and satises limt (s , t ) = 0. A function : R R0 belongs to class PD (is positive denite) if it is continuous and positive everywhere except at the origin.
SOCN 2013
9 / 45
SOCN 2013
10 / 45
Stability Denitions
Suppose X Rn is positive invariant for x + = f (x ), that A is closed and positive invariant for x + = f (x ), and that A lies in the interior of X . Then A is
1
locally stable in X if, for each > 0, there exists a = () > 0 such that x X (A B ), implies |(i ; x )|A < for all i I0 .a locally attractive in X if there exists a > 0 such that x X (A B ) implies |(i ; x )|A 0 as i . attractive in X if |(i ; x )|A 0 as i for all x X .
B denotes the unit ball in Rn .
3 a
SOCN 2013
11 / 45
SOCN 2013
12 / 45
locally asymptotically stable in X if it is locally stable in X and locally attractive in X . asymptotically stable with a region of attraction X if it is locally stable in X and attractive in X . locally exponentially stable with a region of attraction X if there exist > 0, c > 0, and (0, 1) such that x X (A B ) implies |(i ; x )|A c |x |A i for all i I0 . exponentially stable with a region of attraction X if there exists a c > 0 and a (0, 1) such that |(i ; x )|A c |x |A i for all x X , all i I0 .
See Teel and Zaccarian (2006) and the Notes on Recent MPC Literature link on: www.che.wisc.edu/~jbraw/mpc for further discussion of the dierences in the two denitions. If f () is continuous, the two denitions are equivalent.
SOCN 2013 MPC short course 14 / 45
SOCN 2013
13 / 45
Lyapunov function
Denition 7 (Lyapunov function)
A function V : Rn R0 is said to be a Lyapunov function for the system x + = f (x ) and set A if there exist functions i K , i = 1, 2 and 3 PD such that for any x Rn , V (x ) 1 (|x |A ) V (x ) 2 (|x |A ) V (f (x )) V (x ) 3 (|x |A ) (2) (3) (4)
Some ne print
Remark 8
It is shown in Jiang and Wang (2002), Lemma 2.8, that, under the assumption that f () is continuous, we can always assume that 3 () in (4) is a K function.
If V () satises (2)(4) for all x X where X A is a positive invariant set for x + = f (x ), then V () is said to be a Lyapunov function in X for the system x + = f (x ) and set A.
SOCN 2013
15 / 45
SOCN 2013
16 / 45
The existence of a Lyapunov function is a sucient condition for global asymptotic stability as shown in the next result which we prove under the assumption, common in MPC, that 3 () is K function.
SOCN 2013
17 / 45
SOCN 2013
18 / 45
Theorem 12 (Lyapunov function for exponential stability) Theorem 11 (Lyapunov function for asymptotic stability (constrained case))
Suppose X Rn is positive invariant for x + = f (x ), that A is closed and positive invariant for x + = f (x ), and that A lies in the interior of X . If there exists a Lyapunov function in X for the system x + = f (x ) and set A with 3 () a K function, then A is asymptotically stable for x + = f (x ) with a region of attraction X . Suppose X Rn is positive invariant for x + = f (x ), that A is closed and positive invariant for x + = f (x ), and that A lies in the interior of X . If there exists V : Rn R0 satisfying the following properties for all x X
a1 |x | A V (x ) a2 |x |A
SOCN 2013
19 / 45
SOCN 2013
20 / 45
Robust Stability
We turn now to stability conditions for systems subject to bounded disturbances (not vanishingly small) and described by x + = f (x , w ) where the disturbance w lies in the compact set W. This system may equivalently be described by the dierence inclusion x + F (x ) where the set F (x ) := {f (x , w ) | w W} Let S (x ) denote the set of all solutions of (5) or (6) with initial state x . (6) (5)
Assume that the origin is globally exponentially stable (GES) for the system x + = f (x ) in which f is Lipschitz continuous. Show that there exists a Lipschitz continuous Lyapunov function V () for the system satisfying for all x Rn a1 |x | V (x ) a2 |x | V (f (x )) V (x ) a3 |x | in which a1 , a2 , a3 , > 0. Hint: Consider summing the solution |(i ; x )| on i as a candidate Lyapunov function V (x ). Establish also that in the Lyapunov function dened above, any > 0 is valid, and the constant a3 can be chosen as large as one wishes.
SOCN 2013 MPC short course 21 / 45
SOCN 2013
22 / 45
We require, in the sequel, that the set A is positive invariant for (5) (or for x + F (x )):
SOCN 2013
23 / 45
SOCN 2013
24 / 45
Asymptotic stability
Lyapunov function
To cope with disturbances we require a modied denition of a Lyapunov function.
SOCN 2013
25 / 45
SOCN 2013
26 / 45
GAS (disturbances)
Input-to-State Stability
In input-to-state stability (Sontag and Wang, 1995; Jiang and Wang, 2001) we seek a bound on the state in terms of a uniform bound on the disturbance sequence w := {w (0), w (1), . . .}. Let denote the usual norm for sequences, i.e., w := supk 0 |w (k )|.
Inequality 9 ensures V (f (x , w )) V (x ) 3 (|x |A ) for all w W. The existence of a Lyapunov function for the system x + F (x ) and set A is a sucient condition for A to be globally asymptotically stable for x + F (x ) as shown in the next result.
SOCN 2013
27 / 45
SOCN 2013
28 / 45
Feasible sets
The set of feasible initial states and associated control sequences
= f (x , u )
(10) ZN = {(x , u) | u (k ) U, (k ; x , u) X for all k I0:N 1 , and (N ; x , u) Xf } and Xf X is the feasible terminal set. The set of feasible initial states is XN = {x Rn | u UN such that (x , u) ZN } (11)
Given an integer N (referred to as the nite horizon), and an input sequence u of length N , u = {u (0), u (1), . . . , u (N 1)}, let (k ; x , u) denote the solution of (10) at time k for a given initial state x (0) = x . Terminal constraint (and penalty) (N ; x , u) Xf X
SOCN 2013
29 / 45
SOCN 2013
30 / 45
The control law is N (x ) = u 0 (0; x ) The optimum may not be unique; then N () is a point-to-set map Closed-loop system x + = f (x , N (x )) x + f (x , N (x )) dierence equation dierence inclusion
VN (x , u) =
k =0
(x , u ) is the stage cost; Vf (x (N )) is the terminal cost Consider the nite horizon optimal control problem PN (x ) :
uUN
min VN (x , u)
Nominal closed-loop stability question; is the origin stable? If yes, what is the region of attraction? All of XN ?
SOCN 2013
31 / 45
SOCN 2013
32 / 45
Consider a process disturbance d , x + = f (x , (x )) + d A measurement disturbance xm = x + e Nominal controller with disturbance x + f (x , N (xm )) + d x + f (x , N (x + e )) + d x + F (x , w ) w = (d , e )
SOCN 2013
33 / 45
SOCN 2013
34 / 45
0 () is a Lyapunov function for the We show that the optimal cost VN closed-loop system. We require three properties. Lower bound. 0 VN (x ) 1 (|x |)
for all
x XN
Given the denition of VN (x , u) as a sum of stage costs, we have using Assumption 23 VN (x , u) (x , u (0; x )) 1 (|x |) so the rst property is established. for all x XN , u UN
for all
x XN
At state x XN , consider the optimal sequence u0 (x ) = {u (0; x ), u (1; x ), . . . , u (N 1; x )}, and generate a candidate sequence for the successor state, x + := f (x , N (x )) = {u (1; x ), u (2; x ), . . . , u (N 1; x ), f (x (N ))} u with x (N ) := (N ; x , u). This candidate is feasible for x + because Xf is control invariant under control law f () (Assumption 22). The cost is
0 ) = VN VN (x + , u (x ) (x , u (0; x ))
for all
x Xf
for all
x XN
Vf (x (N )) + (x (N ), f (x (N ))) + Vf (f (x (N ), f (x (N ))))
SOCN 2013 MPC short course 37 / 45 SOCN 2013 MPC short course 38 / 45
Upper bound
Finally we require the upper bound.
0 VN (x ) 2 (|x |)
x X2
Surprisingly, this one turns out to be the most involved. First, we have the bound from Assumption 23 Vf (x ) 2 (|x |) Next we show that Consider N = 1,
0 (x ) VN
x X2 x X1 x X1
(x , 1 (x )) + = Therefore
0 V1 (x )
0 V1 (f (x , 1 (x )))
for all
x Xf
(x , 1 (x )) + Vf (f (x , 1 (x ))) x X1
Vf (x ) for x Xf , N 1.
0 V2 (x ) Vf (x ) x Xf u U
0 (x ) = min{ (x , u ) + Vf (f (x , u )) | f (x , u ) Xf } V1
Vf (x )
x Xf
x Xf
SOCN 2013
39 / 45
SOCN 2013
40 / 45
Therefore, the origin is asymptotically stable (KL version) with region of attraction XN . We can also establish robust stability, but lets do that later. If we strengthen the properties of (), we can strengthen the conclusion to exponential stability.
0 () plays in the stability analysis of Notice the essential role that VN MPC.
for all
x XN
Be aware that the MPC literature has been confused about the requirements for this last result.
See Proposition 10 of Notes on Recent MPC Literature link on: www.che.wisc.edu/~jbraw/mpc. Thanks also to Andy Teel.
SOCN 2013 MPC short course 41 / 45 2
In economic MPC we lose this Lyapunov function and have to work to bring it back.
SOCN 2013
42 / 45
Recommended exercises
Further Reading I
H. Chen and F. Allg ower. A quasi-innite horizon nonlinear model predictive control scheme with guaranteed stability. Automatica, 34(10): 12051217, 1998. Z.-P. Jiang and Y. Wang. Input-to-state stability for discrete-time nonlinear systems. Automatica, 37:857869, 2001. Z.-P. Jiang and Y. Wang. A converse Lyapunov theorem for discrete-time systems with disturbances. Sys. Cont. Let., 45:4958, 2002. H. K. Khalil. Nonlinear Systems. Prentice-Hall, Upper Saddle River, NJ, third edition, 2002. J. B. Rawlings and D. Q. Mayne. Model Predictive Control: Theory and Design. Nob Hill Publishing, Madison, WI, 2009. 576 pages, ISBN 978-0-9759377-0-9.
Stability denitions. Example 2.3 Lyapunov functions. Exercise B.2B.4.4 Dynamic programming. Exercise C.1C.2.4 MPC stability results. Theorem 7 and Example 1.3 Exercises 2.11, 2.14, 2.154
Notes on Recent MPC Literature link on: www.che.wisc.edu/~jbraw/mpc. Rawlings and Mayne (2009, Appendices B and C). Downloadable from www.che.wisc.edu/~jbraw/mpc.
4 SOCN 2013 MPC short course 43 / 45
E. D. Sontag and Y. Wang. On the characterization of the input to state stability property. Sys. Cont. Let., 24:351359, 1995.
SOCN 2013 MPC short course 44 / 45
Further Reading II
A. R. Teel and L. Zaccarian. On uniformity in denitions of global asymptotic stability for time-varying nonlinear systems. Automatica, 42: 22192222, 2006.
SOCN 2013
45 / 45