8/9/2019 04 mpc jbr
1/23
Preliminaries for Model Predictive Control course
James B. Rawlings
Department of Chemical and Biological EngineeringUniversity of Wisconsin–Madison
Insitut für Systemtheorie und RegelungstechnikUniversität StuttgartStuttgart, Germany
June 20, 2011
Stuttgart – June 2011 MPC short course 1 / 46
Outline
1 Introduction
2 Stability, equilibria, invariant sets
3 Lyapunov function theory
4 Disturbances and robust stability
5 Basic MPC problem and nominal stability
Stuttgart – June 2011 MPC short course 2 / 46
8/9/2019 04 mpc jbr
2/23
Predictive control
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0 0
0 0
1 1
1 1
0 0
0 0
0 0
1 1
1 1
1 1
0
0
1
1
0
0
1
1
0
0
0
1
1
1
Measurement
MH EstimateMPC control
Forecast
t time
Reconcile the past Forecast the future
sensorsy
actuatorsu
minu (t )
T 0|y sp − g (x , u )|
2Q
+ |u sp − u |2R
dt
ẋ = f (x , u )
x (0) = x 0 (given)
y = g (x , u )Stuttgart – June 2011 MPC short course 3 / 46
State estimation
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0
0
1
1
0 00 01 11 1 0 00 00 01 11 11 1
0
0
1
1
0
0
1
1
000111
Measurement
MH EstimateMPC control
Forecast
t time
Reconcile the past Forecast the future
sensorsy
actuatorsu
minx 0,w (t )
0−T
|y − g (x , u )|2R + |ẋ − f (x , u )|2Q dt
ẋ = f (x , u ) + w (process noise)
y = g (x , u ) + v (measurement noise)
Stuttgart – June 2011 MPC short course 4 / 46
8/9/2019 04 mpc jbr
3/23
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 andcontrol, and x + the successor state.
We assume in the sequel that the function f : Rn × Rm → Rn iscontinuous.
Letφ(k ; x , u)
denote the solution of x + = 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.
1Most of this preliminary material is taken from Rawlings and Mayne (2009,Appendix B). Downloadable from www.che.wisc.edu/~jbraw/mpc.
Stuttgart – June 2011 MPC short course 5 / 46
Existence of solutions to model
If a state-feedback control law u = κ(x ) has been chosen, theclosed-loop system is described by x + = f (x , κ(x )).
Let φ(k ; x , κ(·)) denote the solution of this difference equation attime k if the initial state at time 0 is x (0) = x ; the solution exists andis unique (even if κ(·) is discontinuous).
If κ(·) is not continuous, as may be the case when κ(·) is a model
predictive control (MPC) law, then f ((·), κ(·)) may not be continuous.In this case we assume that f ((·), κ(·)) is locally bounded .
Definition 1 (Locally bounded)
A function f : X → X is locally bounded if, for any x ∈ X , there exists aneighborhood N of x such that f ( N ) is a bounded set, i.e., if there existsa M > 0 such that |f (x )| ≤ M for all x ∈ N .
Stuttgart – June 2011 MPC short course 6 / 46
http://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpc
8/9/2019 04 mpc jbr
4/23
Stability and equilibrium point
We would like to be sure that the controlled system is “stable”, i.e., thatsmall perturbations of the initial state do not cause large variations in thesubsequent behavior of the system, and that the state converges to adesired state or, if this is impossible due to disturbances, to a desired setof states.If convergence to a specified state, x ∗ say, is sought, it is desirable for thisstate to be an equilibrium point:
Definition 2 (Equilibrium point)
A point x ∗ is an equilibrium point of x + = f (x ) if x (0) = x ∗ impliesx (k ) = φ(k ; x ∗) = x ∗ for all k ≥ 0. Hence x ∗ is an equilibrium point if itsatisfies
x ∗ = f (x ∗)
Stuttgart – June 2011 MPC short course 7 / 46
Positive invariant set
In other situations, for example when studying the stability properties of an oscillator, convergence to a specified closed set A ⊂ Rn is sought.If convergence to a set A is sought, it is desirable for the set A to bepositive invariant :
Definition 3 (Positive invariant set)
A set A is positive invariant for the system x + = f (x ) if x ∈ A impliesf (x ) ∈ A.
Clearly, any solution of x + = f (x ) with initial state in A, remains in A.The (closed) set A = {x ∗} consisting of a (single) equilibrium point is aspecial case; x ∈ A (x = x ∗) implies f (x ) ∈ A (f (x ) = x ∗).
Stuttgart – June 2011 MPC short course 8 / 46
8/9/2019 04 mpc jbr
5/23
Distance to a set; set addition
Define distance from point x to set A
|x |A := inf z ∈A|x − z |
If A = {x ∗}, then |x |A = |x − x ∗| which reduces to |x | when x ∗ = 0.
Set addition: A⊕ B := {a + b | a ∈ A, b ∈ B }.
Stuttgart – June 2011 MPC short course 9 / 46
K, K∞, KL, and PD functions
Definition 4
A function σ : R≥0 → R≥0 belongs to class K if it is continuous, zeroat zero, and strictly increasing;
σ : R≥0 → R≥0 belongs to class K∞ if it is a class K and unbounded(σ(s )→∞ as s → ∞).
A function β : R≥0 × I≥0 → R≥0 belongs to class KL if it iscontinuous and if, for each t ≥ 0, β (·, t ) is a class K function and foreach s ≥ 0, β (s , ·) is nonincreasing and satisfies limt →∞ β (s , t ) = 0.
A function γ : R→ R≥0 belongs to class PD (is positive definite) if itis continuous and positive everywhere except at the origin.
Stuttgart – June 2011 MPC short course 10 / 46
8/9/2019 04 mpc jbr
6/23
Some useful properties of K functions
The following useful properties of these functions are established in Khalil(2002, Lemma 4.2):
if α1(·) and α2(·) are K functions (K∞ functions), then α−11 (·) and
(α1 ◦ α2)(·) := α1(α2(·)) are K functions (K∞ functions).
Moreover, if α1(·) and α2(·) are K functions and β (·) is a KLfunction, then σ(r , s ) = α1(β (α2(r ), s )) is a KL function.
Stuttgart – June 2011 MPC short course 11 / 46
Stability — Definitions
Definition 5 (Various forms of stability (constrained))
Suppose X ⊂ Rn is positive invariant for x + = f (x ), that A is closed andpositive invariant for x + = f (x ), and that A lies in the interior of X . ThenA is
1 locally stable in X if, for each ε > 0, there exists a δ = δ (ε) > 0 such
that x ∈ X ∩ (A⊕ δ B ), implies |φ(i ; x )|A < ε for all i ∈ I≥0.a
2 locally attractive in X if there exists a η > 0 such thatx ∈ X ∩ (A⊕ ηB ) implies |φ(i ; x )|A → 0 as i → ∞.
3 attractive in X if |φ(i ; x )|A → 0 as i → ∞ for all x ∈ X .
aB denotes the unit ball in Rn.
Stuttgart – June 2011 MPC short course 12 / 46
8/9/2019 04 mpc jbr
7/23
8/9/2019 04 mpc jbr
8/23
Lyapunov function
Definition 7 (Lyapunov function)
A function V : Rn → R≥0 is said to be a Lyapunov function for the system
x +
= f (x ) and set A if there exist functions αi ∈ K∞, i = 1, 2 andα3 ∈ PD such that for any x ∈ Rn,
V (x ) ≥ α1(|x |A) (2)
V (x ) ≤ α2(|x |A) (3)
V (f (x ))− V (x ) ≤ −α3(|x |A) (4)
If V (·) satisfies (2)–(4) for all x ∈ X where X ⊃ A is a positive invariantset for x + = f (x ), then V (·) is said to be a Lyapunov function in X forthe system x + = f (x ) and set A.
Stuttgart – June 2011 MPC short course 15 / 46
Some fine print
Remark 8
It is shown in Jiang and Wang (2002), Lemma 2.8, that, under theassumption that f (·) is continuous, we can always assume that α
3(·) in (4)
is a K∞ function.
Stuttgart – June 2011 MPC short course 16 / 46
8/9/2019 04 mpc jbr
9/23
Lyapunov function — Asymptotic stability
The existence of a Lyapunov function is a sufficient condition for global
asymptotic stability as shown in the next result which we prove under theassumption, common in MPC, that α3(·) is K∞ function.
Theorem 9 (Lyapunov function and GAS)
Suppose V (·) is a Lyapunov function for x + = f (x ) and set A with α3(·)a K∞ function. Then A is globally asymptotically stable.
Stuttgart – June 2011 MPC short course 17 / 46
Converse Lyapunov theorem — Asymptotic stability
Theorem 9 merely provides a sufficient condition for global asymptoticstability that might be thought to be conservative. The next result, aconverse stability theorem by Jiang and Wang (2002), establishes necessityunder a stronger hypothesis, namely that f (·) is continuous rather thanlocally bounded.
Theorem 10 (Converse theorem for asymptotic stability)
Let A be compact and f (·) continuous. Suppose that the set A is globally asymptotically stable for the system x + = f (x ). Then there exists asmooth Lyapunov function for the system x + = f (x ) and set A.
A proof of this result is given in Jiang and Wang (2002), Theorem 1,part 3.
Stuttgart – June 2011 MPC short course 18 / 46
8/9/2019 04 mpc jbr
10/23
Asymptotic stability — Constrained case
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 Awith α3(·) a K∞ function, then A is asymptotically stable for x + = f (x )with a region of attraction X.
Stuttgart – June 2011 MPC short course 19 / 46
Exponential stability — Constrained case
Theorem 12 (Lyapunov function for exponential stability)
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 → R≥0 satisfying the following properties for all x ∈ X
a1 |x |σA ≤ V (x ) ≤ a2 |x |σAV (f (x ))− V (x ) ≤ −a3 |x |
σ
A
in which a1, a2, a3, σ > 0, then A is exponentially stable for x + = f (x )with a region of attraction X.
Stuttgart – June 2011 MPC short course 20 / 46
8/9/2019 04 mpc jbr
11/23
Converse theorem for exponential stability
Exercise B.3: A converse theorem for exponential stability
a Assume that the origin is globally exponentially stable (GES) for thesystem
x + = f (x )
in which f is Lipschitz continuous. Show that there exists a Lipschitzcontinuous Lyapunov function V (·) for the system satisfying for allx ∈ Rn
a1 |x |σ ≤ V (x ) ≤ a2 |x |
σ
V (f (x ))− V (x ) ≤ −a3 |x |σ
in which a1, a2, a3, σ > 0.Hint: Consider summing the solution |φ(i ; x )| on i as a candidateLyapunov function V (x ).
b Establish also that in the Lyapunov function defined above, any σ > 0is valid, and the constant a3 can be chosen as large as one wishes.
Stuttgart – June 2011 MPC short course 21 / 46
Robust Stability
We turn now to stability conditions for systems subject to boundeddisturbances (not vanishingly small) and described by
x + = f (x , w ) (5)
where the disturbance w lies in the compact set W.This system may equivalently be described by the difference inclusion
x + ∈ F (x ) (6)
where the setF (x ) := {f (x , w ) | w ∈W}
Let S (x ) denote the set of all solutions of (5) or (6) with initial state x .
Stuttgart – June 2011 MPC short course 22 / 46
8/9/2019 04 mpc jbr
12/23
Positive invariant sets
We require, in the sequel, that the set A is positive invariant for (5) (orfor x + ∈ F (x )):
Definition 13 (Positive invariance with disturbances)
The set A is positive invariant for x + = f (x , w ), w ∈W if x ∈ A impliesf (x , w ) ∈ A for all w ∈W; it is positive invariant for x + ∈ F (x ) if x ∈ Aimplies F (x ) ⊆ A.
Clearly the two definitions are equivalent; A is positive invariant forx + = f (x , w ), w ∈W, if and only if it is positive invariant for x + ∈ F (x ).In Definitions 14-16, we use “positive invariant” to denote “positive
invariant for x + = f (x , w ), w ∈W” or for x + ∈ F (x ).
Stuttgart – June 2011 MPC short course 23 / 46
Stability and attraction
Definition 14 (Local stability (disturbances))
The closed positive invariant set A is locally stable for x + = f (x , w ),w ∈W (or for x + ∈ F (x )) if, for all ε > 0, there exists a δ > 0 such that,for each x satisfying |x |A < δ , each solution φ ∈ S (x ) satisfies |φ(i )|A < εfor all i ∈ I≥0.
Definition 15 (Global attraction (disturbances))
The closed positive invariant set A is globally attractive for the systemx + = f (x , w ), w ∈W (or for x + ∈ F (x )) if, for each x ∈ Rn, eachsolution φ(·) ∈ S (x ) satisfies |φ(i )|A → 0 as i → ∞.
Stuttgart – June 2011 MPC short course 24 / 46
8/9/2019 04 mpc jbr
13/23
Asymptotic stability
Definition 16 (GAS (disturbances))
The closed positive invariant set A is globally asymptotically stable forx + = f (x , w ), w ∈W (or for x + ∈ F (x )) if it is locally stable and globallyattractive.
An alternative definition of global asymptotic stability of A forx + = f (x , w ), w ∈W, if A is compact, is the existence of a KL functionβ (·) such that for each x ∈ Rn, each φ ∈ S (x ) satisfies|φ(i )|A ≤ β (|x |A, i ) for all i ∈ I≥0.
Stuttgart – June 2011 MPC short course 25 / 46
Lyapunov function
To cope with disturbances we require a modified definition of a Lyapunovfunction.
Definition 17 (Lyapunov function (disturbances))
A function V : Rn → R≥0 is said to be a Lyapunov function for the systemx + = f (x , w ), w ∈W (or for x + ∈ F (x )) and set A if there exist
functions αi ∈ K∞, i = 1, 2 and α3 ∈ PD such that for any x ∈ Rn
,
V (x ) ≥ α1(|x |A) (7)
V (x ) ≤ α2(|x |A) (8)
supz ∈F (x )
V (z )− V (x ) ≤ −α3(|x |A) (9)
Stuttgart – June 2011 MPC short course 26 / 46
8/9/2019 04 mpc jbr
14/23
GAS (disturbances)
Inequality 9 ensures V (f (x , w ))− V (x ) ≤ −α3(|x |A) for all w ∈W. Theexistence of a Lyapunov function for the system x + ∈ F (x ) and set A is a
sufficient condition for A to be globally asymptotically stable forx + ∈ F (x ) as shown in the next result.
Theorem 18 (Lyapunov function for GAS (disturbances))
Suppose V (·) is a Lyapunov function for x + = f (x , w ), w ∈W (or for x + ∈ F (x )) and set A with α3(·) a K∞ function. Then A is globally asymptotically stable for x + = f (x , w ), w ∈W (or for x + ∈ F (x )).
Stuttgart – June 2011 MPC short course 27 / 46
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 thedisturbance sequence w := {w (0), w (1), . . .}. Let · denote the usual ∞norm for sequences, i.e., w := supk ≥0 |w (k )|.
Definition 19 (Input-to-state stable (ISS))
The system x + = f (x , w ) is (globally) input-to-state stable (ISS) if there
exists a KL function β (·) and a K function σ(·) such that, for eachx ∈ Rn, and each disturbance sequence w = {w (0), w (1), . . .} in ∞
|φ(i ; x , wi )| ≤ β (|x | , i ) + σ(wi )
for all i ∈ I≥0, where φ(i ; x , wi ) is the solution, at time i , if the initial stateis x at time 0 and the input sequence is wi := {w (0), w (1), . . . , w (i − 1)}.
Stuttgart – June 2011 MPC short course 28 / 46
8/9/2019 04 mpc jbr
15/23
The basic nonlinear, constrained MPC problem
The system model isx + = f (x , u ) (10)
Both state and input are subject to constraints
x (k ) ∈ X , u (k ) ∈ U for all k ∈ I≥0
Given an integer N (referred to as the finite horizon), and an inputsequence u of length N , u = {u (0), u (1), . . . , u (N − 1)}, let φ(k ; x , u)denote the solution of (10) at time k for a given initial state x (0) = x .
Terminal constraint (and penalty)
φ(N ; x , u) ∈ Xf ⊆ X
Stuttgart – June 2011 MPC short course 29 / 46
Feasible sets
The set of feasible initial states and associated control sequences
ZN = {(x , u) | u (k ) ∈ U, φ(k ; x , u) ∈ X for all k ∈ I0:N −1,
and φ(N ; x , u) ∈ Xf }
and Xf ⊆ X is the feasible terminal set.
The set of feasible initial states is
X N = {x ∈ Rn | ∃u ∈ UN such that (x , u) ∈ ZN } (11)
For each x ∈ X N , the corresponding set of feasible input sequences is
U N (x ) = {u | (x , u) ∈ ZN }
Stuttgart – June 2011 MPC short course 30 / 46
8/9/2019 04 mpc jbr
16/23
Cost function and control problem
For any state x ∈ Rn and input sequence u ∈ UN , we define
V N (x , u) =N −1k =0
(φ(k ; x , u), u (k )) + V f (φ(N ; x , u))
(x , u ) is the stage cost; V f (x (N )) is the terminal cost
Consider the finite horizon optimal control problem
PN (x ) : minu∈U N
V N (x , u)
Stuttgart – June 2011 MPC short course 31 / 46
Control law and closed-loop system
The control law isκN (x ) = u
0(0; x )
The optimum may not be unique; then κN (·) is a point-to-set map
Closed-loop system
x +
= f (x , κN (x )) difference equation
x + ∈ f (x , κN (x )) difference inclusion
Nominal closed-loop stability question; is the origin stable?
If yes, what is the region of attraction? All of X N ?
Stuttgart – June 2011 MPC short course 32 / 46
8/9/2019 04 mpc jbr
17/23
Inherent robustness of the nominal controller
Consider a process disturbance d , x + = f (x , κ(x )) + d
A measurement disturbance x m = x + e
Nominal controller with disturbance
x + ∈ f (x , κN (x m)) + d
x + ∈ f (x , κN (x + e )) + d
x + ∈ F (x , w ) w = (d , e )
Robust stability; is the system x + ∈ F (x , w ) input-to-state stableconsidering w = (d , e ) as the input.
Stuttgart – June 2011 MPC short course 33 / 46
Basic MPC assumptions
Assumption 20 (Continuity of system and cost)
The functions f : Rn ×Rm → Rn, : Rn × Rm → R≥0 andV f : R
n → R≥0 are continuous, f (0, 0) = 0, (0, 0) = 0, and V f (0) = 0.
Assumption 21 (Properties of constraint sets)
The set U is compact and contains the origin. The sets X and Xf areclosed and contain the origin in their interiors, Xf ⊆ X.
Stuttgart – June 2011 MPC short course 34 / 46
8/9/2019 04 mpc jbr
18/23
Basic MPC assumptions
Assumption 22 (Basic stability assumption)
For any x ∈ Xf there exists u := κf (x ) ∈ U such that f (x , u ) ∈ Xf andV f (f (x , u )) + (x , u ) ≤ V f (x ).
Note: understanding this requirement created a big research challenge forthe development of nonlinear MPC. Credit the celebrated quasi-infinitehorizon work of Chen and Allgöwer (1998) for cracking this problem.
Assumption 23 (Bounds on stage and terminal costs)
The stage cost (·) and the terminal cost V f (·) satisfy
(x , u ) ≥ α1(|x |) ∀x ∈ X N , ∀u ∈ U
V f (x ) ≤ α2(|x |) ∀x ∈ Xf
in which α1(·) and α2(·) are K∞ functions
Stuttgart – June 2011 MPC short course 35 / 46
Optimal MPC cost function as Lyapunov function
We show that the optimal cost V 0N (·) is a Lyapunov function for theclosed-loop system. We require three properties.Lower bound.
V 0N (x ) ≥ α1(|x |) for all x ∈ X N
Given the definition of V N (x , u) as a sum of stage costs, we have usingAssumption 23
V N (x , u) ≥ (x , u (0; x )) ≥ α1(|x |) for all x ∈ X N , u ∈ UN
so the first property is established.
Stuttgart – June 2011 MPC short course 36 / 46
8/9/2019 04 mpc jbr
19/23
MPC cost function as Lyapunov function – cost decrease
Next we require the cost decrease
V 0N (f (x , κN (x )) ≤ V 0N (x )− α3(|x |) for all x ∈ X N
At state x ∈ X N , 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 , κN (x ))
ũ = {u (1; x ), u (2; x ), . . . , u (N − 1; x ), κf (x (N ))}
with x (N ) := φ(N ; x , u). This candidate is feasible for x + because Xf iscontrol invariant under control law κf (·) (Assumption 22).The cost is
V N (x +, ũ) = V 0N (x )− (x , u (0; x ))
− V f (x (N )) + (x (N ), κf (x (N ))) + V f (f (x (N ), κf (x (N ))))
Stuttgart – June 2011 MPC short course 37 / 46
Cost decrease (cont.)
But by Assumption 22
V f (f (x , κf (x ))) + (x , κf (x )) ≤ V f (x ) for all x ∈ Xf
so we have that
V N (x +, ũ) ≤ V 0N (x )− (x , u (0; x ))
The optimal cost is certainly no worse, giving
V 0N (x +) ≤ V 0N (x )− (x , u (0; x ))
V 0N (x +) ≤ V 0N (x )− α1(|x |) for all x ∈ X N
which is the desired cost decrease with the choice α3(·) = α1(·).
Stuttgart – June 2011 MPC short course 38 / 46
8/9/2019 04 mpc jbr
20/23
Upper bound
Finally we require the upper bound.
V 0N (x ) ≤ α2(|x |) for all x ∈ X N
Surprisingly, this one turns out to be the most involved.First, we have the bound from Assumption 23
V f (x ) ≤ α2(|x |) for all x ∈ Xf
Next we show that V 0N (x ) ≤ V f (x ) for x ∈ Xf , N ≥ 1.Consider N = 1,
V 01 (x ) = minu ∈U{(x , u ) + V f (f (x , u )) | f (x , u ) ∈ Xf }
≤ V f (x ) x ∈ Xf
Stuttgart – June 2011 MPC short course 39 / 46
Dynamic programming recursion
Next consider N = 2, and optimal control law κ2(·)
V 02 (x ) = minu ∈U{(x , u ) + V 01 (f (x , u )) | f (x , u ) ∈ X 1} x ∈ X 2
= (x , κ2(x )) + V 0
1 (f (x , κ2(x ))) x ∈ X 2
≤ (x , κ1(x )) + V 0
1 (f (x , κ1(x ))) x ∈ X 1
≤ (x , κ1(x )) + V f (f (x , κ1(x ))) x ∈ X 1
= V 0
1 (x ) x ∈ X 1
ThereforeV 02 (x ) ≤ V f (x ) x ∈ Xf
Continuing this recursion gives for all N ≥ 1
V 0N (x ) ≤ V f (x ) x ∈ Xf
Stuttgart – June 2011 MPC short course 40 / 46
8/9/2019 04 mpc jbr
21/23
Extending the upper bound from Xf to X N
Question: When can we extend this bound from Xf to the (possiblyunbounded!) set X N ? Recall that V
0N (·) is not necessarily continuous.
Answer: A function can be upper bounded by a K∞ function if and
only if it is locally bounded.2
We know from continuity of f (·) (Assumption 20) that V N (x , u) is acontinuous function, hence locally bounded, and therefore so isV 0N (x ).Therefore, there exists β (·) ∈ K∞ such that
V 0N (x ) ≤ β (|x |) for all x ∈ X N
Be aware that the MPC literature has been confused about the
requirements for this last result.
2See Proposition 10 of “Notes on Recent MPC Literature” link on:www.che.wisc.edu/~jbraw/mpc. Thanks also to Andy Teel.
Stuttgart – June 2011 MPC short course 41 / 46
Asymptotic stability of constrained nonlinear MPC
Why you want a Lyapunov function
We have established that the optimal cost V 0N (·) is a Lyapunovfunction on X N for the closed-loop system.
Therefore, the origin is asymptotically stable (KL version) with regionof attraction X N .
We can also establish robust stability, but let’s do that later.
If we strengthen the properties of (·), we can strengthen theconclusion to exponential stability.
Notice the essential role that V 0N (·) plays in the stability analysis of MPC.
In economic MPC we lose this Lyapunov function and have to work tobring it back.
Stuttgart – June 2011 MPC short course 42 / 46
http://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpc
8/9/2019 04 mpc jbr
22/23
Recommended exercises
Stability definitions. Example 2.3
Lyapunov functions. Exercise B.2–B.4.4
Dynamic programming. Exercise C.1–C.2.4
MPC stability results. Theorem 7 and Example 1.3
Exercises 2.11, 2.14, 2.154
3“Notes on Recent MPC Literature” link on: www.che.wisc.edu/~jbraw/mpc.4Rawlings and Mayne (2009, Appendices B and C). Downloadable from
www.che.wisc.edu/~jbraw/mpc.Stuttgart – June 2011 MPC short course 43 / 46
Further Reading I
H. Chen and F. Allgöwer. A quasi-infinite horizon nonlinear modelpredictive control scheme with guaranteed stability. Automatica, 34(10):1205–1217, 1998.
Z.-P. Jiang and Y. Wang. Input-to-state stability for discrete-timenonlinear systems. Automatica, 37:857–869, 2001.
Z.-P. Jiang and Y. Wang. A converse Lyapunov theorem for discrete-timesystems with disturbances. Sys. Cont. Let., 45:49–58, 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, ISBN978-0-9759377-0-9.
E. D. Sontag and Y. Wang. On the characterization of the input to statestability property. Sys. Cont. Let., 24:351–359, 1995.
Stuttgart – June 2011 MPC short course 44 / 46
http://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpchttp://www.che.wisc.edu/~jbraw/mpc
8/9/2019 04 mpc jbr
23/23
Further Reading II
A. R. Teel and L. Zaccarian. On “uniformity” in definitions of globalasymptotic stability for time-varying nonlinear systems. Automatica, 42:2219–2222, 2006.
Stuttgart – June 2011 MPC short course 45 / 46
Acknowledgments
The author is indebted to Luo Ji of the University of Wisconsin andGanzhou Wang of the Universität Stuttgart for their help in organizing thematerial and preparing the overheads.