Vous êtes sur la page 1sur 20

Gain-scheduling MPC of nonlinear systems

L. Chisci, P. Falugi and G. Zappa


L. Chisci, P. Falugi and G. Zappa,
Dipartimento di Sistemi e Informatica Universit`a di Firenze,
via Santa Marta 3, 50139 Firenze. Italy
e-mail: chisci,falugi,zappa@dsi.uni.it
Abstract
Predictive control of nonlinear systems is addressed by embedding
the dynamics into an LPV system and by computing robust invariant
sets. This mitigates the on-line computational burden by transferring
most of the computations o-line. Benets and conservatism of this
approach are discussed in relation with the control of a critical me-
chanical system.
keywords: Predictive control, constraints, nonlinear systems, LPV sys-
tems, invariant sets, gain-scheduling control.
1 INTRODUCTION
Model Predictive Control (MPC) of nonlinear systems has recently attracted
great interest in the control literature [1]. In particular, there exist several
receding-horizon control schemes [2]-[7] which successfully address the issues
of stability, constraint satisfaction and performance optimization for nonlin-
ear systems. The common idea of these approaches is to enforce stability
by introducing in the control optimization problem suitable terminal con-
straints and penalties derived from an auxiliary, usually linear, control law.
Despite of the remarkable performance and nice theoretical properties of
such predictive controllers, there is a major computational drawback associ-
ated with the non-convexity of the optimization problem which, in turn, is
a consequence of the system nonlinearity. Recently, [9] have advocated the
use of linear embedding techniques [8] in the context of nonlinear MPC by
virtue of their computational advantages over direct nonlinear MPC meth-
ods. In particular the algorithm in [9] relies on the embedding of a nonlinear
1
system into a polytopic Linear Time Varying (LTV) uncertain system and
requires a convex LMI optimization to be solved on-line. In this paper,
we adopt a less conservative embedding into an LPV (Linear Parameter
Varying) model [10] and derive an MPC algorithm which involves a simpler
convex Quadratic Programming (QP) problem to be solved on-line. LPV
models have already been employed successfully in the context of robust lin-
ear MPC [11, 12]; here they are considered in the context of nonlinear MPC.
Following the paradigm, frequently adopted in gain-scheduling control de-
sign [13, 14], of embedding the original nonlinear system into an LPV model,
it is possible, by means of the techniques in [15], to design a gain-scheduling
feedback which stabilizes the system and also to compute, under certain
conditions, an associated admissible set. Within this set of simple structure,
the gain-scheduling feedback guarantees asymptotic stability and constraint
satisfaction. In these computations the key point is the decoupling of the
parameter dynamics from the state dynamics. Subsequently, following the
closed-loop MPC paradigm [16] and adopting the gain-scheduling policy as
auxiliary control law, we propose a novel MPC algorithm for nonlinear sys-
tems based on a robust MPC approach to LPV systems previously presented
in [11]. It is proved that the proposed algorithm, named Gain Scheduling-
MPC (GS-MPC), guarantees asymptotic regulation (at the origin) of the
constrained nonlinear system, and its domain of attraction is evaluated. Sta-
bilization of nonlinear systems under input constraints has been extensively
studied in the literature [17]-[20] from a dierent control design perspective
based on control Lyapunov functions (CLFs). As opposed to MPC, the
CLF approach does not involve on-line computations but suers from two
main drawbacks: (1) cannot deal with state constraints; and (2) requires
the construction of a CLF, which is not an easy task in general.
2 LPV embedding of nonlinear systems
In this section we shall address the embedding of a nonlinear dynamics into
an LPV representation [14]. The LPV model will be subsequently adopted
for designing MPC algorithms. Consider the discrete-time nonlinear system
x(t + 1) = f(x(t), u(t)) (1)
where u(t) IR
m
and x(t) IR
n
are respectively the state and control input
at sample time t, and f(0, 0) = 0. Moreover x(t) is available for feedback
and the system is subject to the constraints
u(t) U, x(t) X (2)
2
where U is a convex polytope and X a polyhedron, both containing the
origin in the interior. The control objective is to regulate the system at the
origin while satisfying the constraints (2). It will be convenient (even if not
necessary) for the subsequent developments to transform the system (1) into
the following quasi-linear form:
x(t + 1) = A(x
1
(t)) x(t) + B(x
1
(t)) v(t) (3)
for a suitable partitioning x = [x

1
, x

2
]

of the state vector and possibly with


the aid of a suitable pre-compensation feedback
u(t) = g(x(t), v(t)). (4)
Remark 1 (Quasi-linearization) - It is clear that quasi-linear repre-
sentations like (3), if they exist, are highly non-unique. Since the subsequent
step will be to embed (3) into a Linear Dierence Inclusion (LDI), a sensible
criterion to be pursued in order to reduce conservatism of the embedding is
to make (3) as simple as possible in the sense that: (1) x
1
has low dimension;
(2) few entries of A(x
1
) and B(x
1
) depend on x
1
. To this end a feedback
pre-compensation (4) can be exploited in order to cancel part of the nonlin-
earities prior to embedding. On the other hand, feedback pre-compensation
may require large control eorts to cancel nonlinearities and may, in turn,
imply conservatism due to the presence of constraints on u. Hence, the ex-
tent to which combine feedback linearization [22] and embedding techniques
is very much system-dependent and a systematic procedure to transform (1)
into (3) is impossible to give. However, it must be pointed out that it is
not convenient, in general, to fully exploit feedback linearization due to its
lack of robustness, but it can be useful to cancel some precisely known non-
linearities so as to simplify the embedding procedure. As will be seen later
the quasi-linearization step is optional in that the embedding techniques de-
scribed in the sequel can directly be applied to the nonlinear dynamics (1).
Then, as customary in gain-scheduling control design [13, 14], we impose
the rate constraint
|x
1
(t + 1) x
1
(t)| , (5)
i.e. the control signal v must guarantee that the time evolution of the x
1
component is suciently slow. Adopting an LPV paradigm, the dynamics
(3) can be embedded into
x(t + 1) = A(p(t)) x(t) +B(p(t)) v(t) (6)
3
where p(t) is a measurable, time-varying, parameter satisfying
|p(t + 1) p(t)| (7)
and the control signal v, according to (4), is constrained by
g(x(t), v(t)) U (8)
and, from (5), must guarantee
|A
1
(p(t))x(t) +B
1
(p(t))v(t) x
1
(t)| (9)
where A
1
, B
1
are the upper submatrices of A, B corresponding to the parti-
tioning x = [x

1
, x

2
]

. Next, for computational reasons, it is assumed, without


loss of generality, that p(t) evolves in a compact set P which has a nite
partition
P =

_
i=1
P
i
. (10)
For each subset P
i
, the LPV dynamics (6) is embedded into a polytopic
dynamics
x(t + 1) F(i)
_
x(t)

v(t)

(11)
where F(i) IR
n(n+m)
is a polytope of matrices [A, B] such that [A(p), B(p)]
F(i), p P
i
. Hereafter the time varying index i(t) will denote the subset
P
i
to which p(t) belongs. Accordingly, the bound (7) on the parameter vari-
ations is transformed into the dierence inclusion i(t + 1) Q(i(t)) where
Q(i) is a suitable subset of the index set I

= {1, 2, . . . , }. Subsequently,
the velocity constraint (9) is transformed into

F
1
(i)
_
x(t)

v(t)

x
1
(t)

< (12)
where F
1
(i) is the polytope of submatrices [A
1
, B
1
] of the matrices [A, B]
F(i). Finally the nonlinear control constraints (8) are approximated by
linear constraints
Lx(t) +Mv(t) 1 (13)
where L and M are matrices of appropriate dimensions. Notice that each
row of L, M denes a scalar constraint. Hence the polytope of vectors [x

, v

satisfying (13) is an inner approximation of the set of vectors satisfying (8).


To summarize, the original nonlinear dynamics (1) has been embedded into
the LPV dynamics
_
i(t + 1) Q(i(t))
x(t + 1) F(i(t)) [x(t)

v(t)

(14)
4
Further (14) is subject to the rate constraints (12), the control constraints
(13) and state constraints x(t) X which, combined together, take the form
L
i(t)
x(t) +M
i(t)
v(t) 1 (15)
for suitable matrices L
i
and M
i
depending on the parameter index i.
Remark 2 (Direct LPV embedding of the nonlinear dynamics)
- It is worth pointing out that the LPV dynamics (14) can directly be
obtained from (1) without need of the intermediate quasi-linear description
(3) and, hence, of the pre-compensation (4). In fact assume that f(x, u) is
C
1
over the domain X U and let J(x, u)

=
_
f
x
,
f
u
_
IR
n(n+m)
denote
the Jacobian matrix. Then it has been shown (see e.g. [23, p.55]) that the
trajectories of the nonlinear system can be embedded in the trajectories of
the LDI (14) provided that
F(i) = Co
i
,
i
= {J(x, u) : (x, u) X
i
U} IR
n(n+m)
where X
i

= {x X : x
1
P
i
} and Co denotes the convex hull. Clearly this
approach does not require feedback pre-compensation. On the other hand,
if J(x, u) depends on several state and/or input variables, the embedding
can be very complicated and/or conservative. For this reason, the quasi-
linearization (3) can be benecial.
Remark 3 (LPV embedding vs. LTV embedding) - Embedding
techniques have been rst considered in the context of nonlinear MPC in [9]
where the nonlinear dynamics x(t+1) = f(x(t), u(t)) is approximated with a
single polytopic LTV model x(t +1) F [x(t)

u(t)

Here, we approximate
the same dynamics with the LPV model (14) which consists of a bank of
polytopic LTV models F(i) indexed by the discrete parameter i, along with
a set-valued model i(t + 1) Q(i(t)) of the discrete parameter dynamics.
Clearly the adoption of multiple local embeddings should yield much better
approximation of the nonlinear dynamics with respect to a single global em-
bedding. The optional quasi-linearization step can give a further reduction
of conservatism with respect to the approach in [9]
Remark 4 (Choice of the parameter rate bound ) - The bound
in (7) is dictated by the speed of variation of the state variable x
1
, chosen
as parameter in the LPV representation. A large clearly implies a more
conservative LPV embedding so that performance may deteriorate due to
5
lack of information on the real plant dynamics. On the other hand, a small
can imply tighter constraints which, in turn, may cause infeasibility. A lim-
itation on the rate of variation |x
1
(t +1) x
1
(t)| may arise from constraints
on x and u. From a designer point of view, a good choice of should be as
close as possible to such a limitation in order to avoid both unnecessarily
stringent constraints and an unnecessarily conservative LPV embedding.
3 Invariant sets for LPV systems
This section discusses the properties of invariant sets in the context of LPV
systems [14]. In particular we shall refer to the LPV system (14) subject to
the linear constraints (15). Recall that the discrete parameter i(t) is actually
related to the state x(t), i.e. i(t) = i(x
1
(t)). Hence, according to (10), the
state space is partitioned as X =
i
X
i
. Moreover, the dynamics (11) can
be regarded as piecewise polytopic, i.e.
x(t + 1) F(i)
_
x(t)

v(t)

if x(t) X
i
. (16)
Nevertheless, for ease of computation of invariant sets, this relationship will
be disregarded and the extended state s(t)

= [i(t)

x(t)

I IR
n
will
be considered hereafter assuming that i(t) evolves independently of x(t).
The relation between x(t) and i(t) will be reconsidered only at the end of
set computations. Notice that a set I IR
n
in the extended state space
is just an l-tuple {
1
,
2
, . . . ,
l
} of subsets of IR
n
.
Assumption 1 - For any value i I of the discrete parameter, there
exists a linear feedback control law v(t) = F
i
x(t) such that the closed-loop
polytopic system x(t + 1) F(i) [x(t)

(F
i
x(t))

is exponentially stable.
A necessary condition for the existence of such a feedback gain F
i
is
clearly exponential stability of the non linear system (1) around the origin. A
sucient condition is quadratic stabilizability [24, 25] of the vertices of F(i).
Such a condition is, however, not necessary. One could use gains F
i
that
are not quadratically (but, for instance, poly-quadratically, polytopically or
polynomially) stabilizing. Notice that Assumption 1 can be made easier
to satisfy by considering a ner partitioning of the parameter space at the
expense of a larger number of polytopic models. The key issue is whether
the linear gain-scheduling control law
v(t) = F
i(t)
x(t) (17)
6
stabilizes the LPV system (14) and satises the constraints (15) in a suitable
neighborhood of the origin, as in such a case the same stability and constraint
satisfaction requirements are also guaranteed for the original nonlinear sys-
tem (1). In order to use (17) as auxiliary control law [7] in an MPC scheme,
it is necessary to determine a, possibly maximal, admissible set [26, 27]
associated to the resulting closed-loop dynamics
_
i(t + 1) Q(i(t))
x(t + 1) F(i(t))
_
x(t)

(F
i(t)
x(t))

(18)
Denition 1 - Given , 0 < 1, a set I IR
n
of the extended
state space is contractive under the closed-loop dynamics (18) if
s(t) =
_
i(t + 1)

1
x(t + 1)
_

In particular if = 1, is called invariant. If, in addition, all states
s = [i, x

satisfy the constraints (15), is called -admissible (ad-


missible if = 1).
Next we would like to nd, for a xed (0, 1], the largest -admissible
set
0
= {
1
0
,
2
0
, . . . ,
l
0
}. Let X
i
c
= {x : (L
i
+ M
i
F
i
)x 1} denote the
set of states which satisfy the constraints (15) under the linear feedback
v = F
i
x, and let X
c

=
_
X
1
c
, X
2
c
, . . . , X

c
_
. Then
i
0
, i I, are computed
with the following set recursion:
O
i
0
= X
i
c
i = 1, 2, . . . ,
O
i
k
= X
i
c

_
_
_
x : F(i)
_
x
F
i
x
_

jQ(i)
O
j
k1
,
_
_
_
k = 1, 2, . . .
i = 1, 2, . . . ,
(19)
For = 1, O
i
k
represents the set of initial states x(0) for which, under the
closed-loop dynamics (18), the constraints (15) are satised up to time k
assuming that i(0) = i. For < 1, O
i
k
is just a subset of the previous one.
Since O
i
k
O
i
k1
, O
i
k
converges to O
i

. Let O

= {O
1

, . . . , O

}, then the
following result holds.
Theorem 1 [15, 28, 29] - O

is the maximal admissible set (i.e.


contractive subset of X
c
). Moreover the system (18) is exponentially sta-
ble with exponential convergence rate < 1 if and only if the set O

7
provided by (19) has non-empty interior.
Notice that, since the uncertain dynamics F(i) is polytopic and the con-
straints (15) are linear, all the sets O
i
k
are described by a nite number of
linear inequalities. Therefore it is important to ascertain whether the set
O

is nitely determined, i.e. there exists a nite k

such that O
k
= O

.
In this respect, the following result can be proved.
Theorem 2 - Assume that the system (18) is exponentially stable with
convergence rate (0, 1) and let . Then if O
k
is bounded for some
k, O

is nitely determined.
Proof: see the Appendix.
Let be a contractive factor which guarantees that O

has non empty


interior and set
0
= O

.
0
represents a set of initial extended states
for which the plant state is asymptotically steered to the origin under the
gain-scheduling linear feedback (17) without violating the constraints. More
precisely
0
is a collection of convex sets
i
0
IR
n
, 1 i , such that if
x(0)
i
0
when i(0) = i, then x(t) 0 without violating constraints. Since
i(0) is related to x
1
(0), it turns out that x(0) will be driven to the origin if
x(0) X
i

i
0
for some i I. Hence the relevant invariant set in the plant
state space IR
n
is

i=1
_
X
i

i
0
_
. The set

0
, union of convex
sets, is invariant under the auxiliary gain-scheduling control law (17). How-
ever, since additional uncertainty is inherently introduced in the extended
dynamics (14) with respect to (16),

0
is not the largest admissible set for
(16). This set has a much more complex structure and will not be consid-
ered. Conversely we shall investigate whether it is possible to enlarge

0
by
making use of a nonlinear feedback v(t) = h(x(t)). Once again, for compu-
tational reasons, we shall consider the LPV dynamics (14) and the extended
state s(t). This motivates the following denition of controlled invariant set.
Denition 2 - is a controlled invariant set for the LPV constrained
system (14)-(15) if for any s = [i, x

, there exists v, depending on x


and i, such that
(i)
_
Q(i)
F(i) [x

_
(controlled invariance)
(ii) L
i
x +M
i
v 1 (constraint satisfaction)
8
Special controlled invariant sets are the sets of all initial states s(0)
which can be steered into
0
by an N-steps feedback control sequence
{v(0), v(1), . . . , v(N 1)}, where each v(k) is allowed to depend on the
current state x(k) and parameter i(k), and such that x(k) and v(k) satisfy
(15). Notice, however, that given a feasible s(0) it is not possible, in general,
to pre-compute at time t = 0 the control sequence {v(0), v(1), . . . , v(N1)}
(open-loop control) which robustly steers s(0) into
0
in N steps. For
receding-horizon operation it is convenient to resort to the so called closed-
loop paradigm [16] by considering a semi-feedback control sequence
v(k) = F
i(k)
x(k) +c(k) k = 0, 1, . . . , N 1 (20)
which combines the auxiliary control law (17) with an open-loop correction
component c. In order to dene the MPC algorithm that will be presented in
the next section, it is necessary to compute o-line the set of initial extended
states s(0), and the relative correction sequences {c(0), c(1), . . . , c(N 1)},
such that s(N) is steered to
0
by the semi-feedback control sequence
(20). This set, denoted by S
N
I IR
n+mN
and represented by S
N
=
{S
1
N
, S
2
N
, . . . , S

N
}, is obtained recursively as follows:
S
i
0
=
i
0
i = 1, 2, . . . ,
S
i
k
=
_

_
_

_
x
c
z
_

_ :
_

_
F(i)
_
x
F
i
x +c
_
z
_

jQ(i)
S
j
k1
, (L
i
+M
i
F
i
)x 1
_

_
k = 1, 2, . . .
i = 1, 2, . . . ,
(21)
At each stage k, c = c(0) and z = [c

(1), c

(2), . . . , c

(k 1)]

provide a valid
k-steps open-loop c-sequence that steers the extended state s(0) = [i, x

to
s(k)
0
. Also notice that, by the above mechanism, the sequence tail z
will also be valid at the subsequent time step to steer the future state to

0
in k 1 steps. The set of extended states s which can be steered to

0
by an N-steps semi-feedback control sequence of the form (20), denoted
hereafter by
N
= {
1
N
,
2
N
, . . . ,

N
}, is clearly the projection of S
N
onto
I IR
n
. In turn this implies that the set of initial states x which can steered
to

0
is

i=1
_
X
i

i
N
_
. Again, since the sets of matrices F(i) are
polytopic, the sets S
i
N
and
i
N
are polytopic as well; hence,

N
is an union
of polytopic sets.
9
4 MPC algorithm
Let us assume that:
(i) the gain-scheduling feedback (17) has been designed o-line so as to sta-
bilize the LPV model (14);
(ii) the admissible set
0
has been computed o-line by iterating (19) until
O
k
= O
k+1
;
(iii) for a given control horizon N, the set S
N
has been computed o-line
via (21).
Then, MPC can be exploited in order to improve performance of the auxil-
iary gain-scheduling controller by means of the degrees of freedom c(k) and
the receding-horizon control strategy. To this end, the following algorithm
is introduced.
Gain-Scheduling MPC (GS-MPC) algorithm. - At each sample
time t, given the state x(t) = [x

1
(t), x

2
(t)]

, let i(t) = i(x


1
(t)) be the index
such that x
1
(t) X
i
and
c(t) = arg min
c(t)
c(t)
2
subject to
_
x(t)
c(t)
_
S
i(t)
N
. (22)
Then apply to the plant the control signal
u(t) = g(x(t), v(t)), v(t) = F
i(t)
x(t) +c(t|t) (23)
where c

(t) = [c

(t|t), c

(t + 1|t), . . . , c

(t +N 1|t)].
The above algorithm selects, at time t, among all admissible sequences c(t)
the one with minimum
2
norm. Since S
i
N
, 1 i , are convex polytopes,
(22)-(23) amounts to a QP problem.
Remark 6 (Performance criterion) - Recalling the denition of
the correction sequence c() in (20), the performance criterion minimized
in (22) is J =

k=0
v(t + k|t) F
i(x(t+k|t))
x(t + k|t)
2
where the sum-
mation is actually truncated at k = N 1 since the auxiliary control
law v(t + k|t) = F
i(x(t+k|t))
x(t + k|t) is hypothesized for k N. Notice
that the above cost penalizes the deviation between the actual control se-
quence and the one provided by the auxiliary gain-scheduling controller.
The choice of this optimization policy only aims at avoiding a cumbersome
10
non-convex optimization problem. From a performance point of view, a cri-
terion which takes into account the mismatch on closed-loop behaviour like
J =

k=0
x(t+k|t
2
Q
+v(t+k|t)F
i(x(t+k|t))
x(t+k|t)
2
R
, x
Q

= x

Qx,
or
J =

k=0
x(t +k|t)
2
Q
+u(t +k|t)
2
R
(24)
would be more appropriate. In such cases, however, the optimization would
be non-convex due to the highly nonlinear and complicated dependence of
x(t +k|t) and u(t +k|t) from the degrees of freedom c(t +k|t).
As far as stability is concerned, the following result holds.
Theorem 3 - Provided that x(0)

N
, the receding-horizon con-
trol (22)-(23) guarantees that: (i) the constraints (15) are satised and (ii)
lim
t
x(t) = 0.
Proof: see the Appendix.
The GS-MPC algorithm ensures, therefore, asymptotic stability to the
origin with domain of attraction

N
.
Remark 7 (Nonlinear MPC vs. min-norm pointwise control)
- Stabilization of constrained nonlinear systems can also be tackled via con-
trol Lyapunov functions (CLF) [17]-[20]. Compared to the CLF approach,
our GS-MPC algorithm: (1) can deal with state (in addition to control)
constraints; (2) does not require determination of a CLF. As a matter of
fact, the o-line determination of the control invariant set
N
discussed in
section 3, provides in a systematic way a polytopic CLF for the nonlinear
system (1). In fact, let

(x)

= max{ : x } denote the norm induced
by the polytope and assume that the set S
N
in (21) has been computed
with a contraction factor (0, 1). Then V (x) =

i(x)
N
(x), induced by the
-contractive controlled invariant polytope
i(x)
N
, turns out to be a CLF for
the plant (1) which can be used in the pointwise min-norm controller [19]
dened as follows
u(x) = arg min
u
_
u
2
| V (f(x, u)) < V (x)
_
.
Another advantage of using MPC is the possibility of optimizing a perfor-
mance criterion. The peculiarity of the criterion adopted in GS-MPC for
11
the sake of computational simplicity, does not take it for granted that the
real system performance under GS-MPC is better than under pointwise min-
norm control. In GS-MPC, however, optimization is based on predictions
several steps-ahead while pointwise min-norm control uses one-step ahead
optimization which may lead to myopic choices. In this respect, GS-MPC
seems to provide a good tradeo between computational and performance
requirements. This fact will be conrmed by the simulation results reported
in the next section.
5 Simulation results
In this section the previously introduced LPV embedding and MPC tech-
niques are applied to the highly nonlinear model of the Furuta pendulum
[21]. The Furuta pendulum is attached to an arm, driven by a dc-motor,
rotating in the horizontal plane. The control objective is to balance the
pendulum in the upward vertical position. The state space model, derived
by the Lagrange method, is
x
2
= x
3
J(x
2
)
_
x
1
x
3
_
+C(x)
_
x
1
x
3
_
=
_
t
3
u
t
6
sin(x
2
)
_
J(x
2
) =
_
1 +t
7
sin
2
(x
2
) t
1
cos(x
2
)
t
4
cos(x
2
) 1
_
C(x) =
_
t
2
+
1
2
t
7
x
2
sin(2x
2
) t
1
x
3
sin(x
2
) +
1
2
t
7
x
1
sin(2x
2
)

1
2
t
8
x
1
sin(2x
2
) t
5
_
(25)
where: x
1
= is the angular speed of the rotating arm; x
2
= and x
3
=

are the angular position and, respectively, velocity of the pendulum ( = 0 is


the vertical upward position); u is the input voltage of the dc-motor subject
to the actuator limits |u| u
max
; t
i
, 1 i 8, are suitable coecients
depending on the physical parameters of the system. Via a suitable pre-
compensation u = g(x, v), it is possible to transform the dynamics into the
simpler form
x
1
= v x
2
= x
3
x = f
3
(x, v)
In fact, this is achieved by choosing g(x, v) =
1
t
3
[g
1
(x) +g
2
(x
2
)v], g
1
(x) =
t
1
t
6
cos(x
2
)sin(x
2
)
1
2
t
1
t
8
x
2
1
sin(2x
2
)cos(x
2
)+t
1
t
5
x
3
cos(x
2
)+t
2
x
1
+t
7
x
1
x
3
sin(2x
2
)+
t
1
x
2
3
sin(x
2
) e g
2
(x
2
) = 1 t
1
t
4
+(t
7
+t
1
t
4
)sin
2
(x
2
) from which it turns out
that f
3
(x, v) = t
6
sin(x
2
)+t
4
cos(x
2
)v+
1
2
t
8
x
2
1
sin(2x
2
)t
5
x
3
Next we applied
12
Euler discretization with sampling period T
s
in order to get a discrete-time
model
x(t + 1) = A(x
1
, x
2
)x(t) +B(x
2
)v(t)
A(x
1
, x
2
) =
_

_
1 0 0
0 1 +T
s
0
0 a(x
1
, x
2
) 1 T
s
t
5
_

_ B(x
2
) =
_

_
T
s
0
b(x
2
)
_

_
a(x
1
, x
2
) = T
s
_
t
6
+t
8
x
2
1
cos(x
2
)
_ sin(x
2
)
x
2
b(x
2
) = T
s
t
4
cos(x
2
).
(26)
As a nal step, p = x
1
has been selected as gain-scheduling parameter.
Further, in order to embed (26) into a family of polytopic systems, we also
assumed bounds |x
2
| x
2,max
and |x
3
| x
3,max
on the other two state
variables. These bounds along with the bounds p = x
1
P
i
, in turn, imply
bounds on the entries a(x
1
, x
2
) and b(x
2
) in (26) namely
a
i,min
a(x
1
, x
2
) a
i,max
for x
1
= p P
i
, b
min
b(x
2
) b
max
from which it is possible to obtain a polytopic inclusion F(i) for each discrete
parameter i as the convex hull of the four matrix vertices characterized by
a
i,min
, a
i,max
, b
min
, b
max
. Further, the following set-valued map Q(i) has
been assumed for the time evolution of the discrete parameter i:
Q(i) = {max{1, i 1}, i, min{i + 1, l}}, for 1 i l.
The control constraints |g(x(t), v(t))| u
max
have been approximated by
v
i(t)
min
v(t) v
i(t)
max
where the bounds v
i
min
, v
i
max
have been obtained con-
sidering the bounds x
1
P
i
as well as |x
2
| x
2,max
and |x
3
| x
3,max
.
The numeric values of all the model parameters are reported in Table 1. A
comment on the specic choice of the parameter bound rate is in order.
Since the parameter rate constraints (9), in this case, yield |v(t)| /T
s
we
selected = T
s
v, v = max
iI
max
_
|v
i
min
|, |v
i
max
|
_
= 100 in order to avoid
tightening of constraints.
The auxiliary gain-scheduling controller has been designed o-line using
the LMI-based design procedure of [25]; the various gains F
i
are reported
in Table 2. Fig. 1 shows the projections of the 3-dimensional sets

i
0
,
1 i 11, on the coordinate planes. A control horizon N = 3 has been
selected and the corresponding polytopes S
i
3
have been computed. To give an
idea of the complexity of such polytopes, the number of linear inequalities
is also reported in Table 2. It is worth to point out, however, that the
computational load required for the solution of QP is quite insensitive to
13
Table 1: Model parameters
model coecients t
1
= 0.0265 t
2
= 1.0524 t
3
= 0.4549 t
4
= 0.6777 t
5
= 0.6058
t
6
= 48.5267 t
7
= 0.0304 t
8
= 0.7776
sampling period T
s
= 0.01
control bound u
max
= 200
state bounds x
1,max
= 10 x
2,max
=
2
9
x
3,max
= 10
parameter rate bound = 1
no. of parameter subranges = 11
parameter subranges P
i
= [x
1,max
+ 2
x
1,max

(i 1), x
1,max
+ 2
x
1,max

i], i = 1, ,
the number of constraints if interior point algorithms [30] are used. Fig.
1 shows the behaviour of GS-MPC starting from an initial state x(0) =
[10, 0.11, 3.44]

which turns out to be infeasible for the gain-scheduling


controller. It can be seen that GS-MPC guarantees asymptotic stability
and constraint satisfaction. We compared our GS-MPC algorithm, based on
embedding, with the direct Nonlinear MPC (NMPC) algorithm of Magni et
al. [7]. For the sake of comparison, the cost adopted for NMPC is (24) where
the weight matrices Q and R have been selected by LQ inverse optimality
arguments so as to give (for the unconstrained model linearized around the
origin) the same feedback gain F
6
(see Table 2) associated to the subset X
6
of IR
3
containing the origin. Fig. 2 compares the behaviours of GS-MPC
and NMPC. We found that the number of ops required by NMPC was
about 10
3
times the op count of GS-MPC. On the other hand notice that,
despite of the much higher on-line computational load of NMPC, the two
algorithms give a similar response.
6 Conclusions
In this paper a novel MPC algorithm for nonlinear systems has been pro-
posed. It is based on the embedding of the nonlinear dynamics into an LPV
system and the construction of (controlled) invariant sets with a simple poly-
topic structure. The advantages of the approach lie in transferring most of
the computations o-line. In fact, at each sampling instant, only a QP
problem must be solved, while the computations of invariant polytopic sets,
carried out o-line, may require a remarkable time. As many other MPC
schemes, this approach exploits an auxiliary control law within a safe ad-
14
Table 2: Gain-scheduling auxiliary feedback and number of constraints
i F
i
n

of constraints
1 F
1
= [0.0323 0.2159 2.0086]10
3
1945
2 F
2
= [0.0310 0.2079 1.7448]10
3
2215
3 F
3
= [0.0314 0.2053 1.5633]10
3
2201
4 F
4
= [0.0309 0.2016 1.4040]10
3
2219
5 F
5
= [0.0307 0.1985 1.3104]10
3
2530
6 F
6
= [0.0310 0.2002 1.2935]10
3
2621
7 F
7
= [0.0313 0.2026 1.3468]10
3
2668
8 F
8
= [0.0313 0.2017 1.4376]10
3
2404
9 F
9
= [0.0321 0.2058 1.5957]10
3
2465
10 F
10
= [0.0331 0.2133 1.8563]10
3
2358
11 F
11
= [0.0351 0.2226 2.1732]10
3
2041
10 5 0 5 10
10
5
0
5
10
x
1
0.2 0 0.2 0.4
6
4
2
0
2
4
x
2
x
3
10 5 0 5 10
0.6
0.4
0.2
0
0.2
0.4
0.6
x
1
x
3
x
2
Figure 1:

0
(solid), state trajectory (solid-star)
15
0 0.5 1 1.5
10
8
6
4
2
0
Arm Speed, x
1
0 0.5 1 1.5
0.2
0.15
0.1
0.05
0
0.05
0.1
0.15
Pendulum Angle, x
2
0 0.5 1 1.5
4
3
2
1
0
1
Pendulum speed, x
3
0 0.5 1 1.5
50
0
50
100
150
200
Control input, u
Figure 2: Time domain responses under the GS-MPC (solid) and NMPC
(dashed) algorithms
missible set and an optimization-based controller within a larger controlled
invariant set. In particular, the cost functional penalizes only the dierence
between the actual control sequence and the control sequence provided by
the auxiliary gain-scheduling controller. Finally it should be emphasized
that this MPC scheme could naturally encompass additive bounded distur-
bances and parametric uncertainties.
References
[1] Allgower F, Zheng A (Eds.). Nonlinear Model Predictive Control
2000, Progress in Systems and Control Theory 26, Birkhauser, Basel,
Switzerland.
[2] Mayne DQ, Michalska H. Receding-horizon control of nonlinear sys-
tems, IEEE Transactions on Automatic Control 1990; 35, 814-824.
[3] Parisini T, Zoppoli R. A receding-horizon regulator for nonlinear sys-
tems and a neural approximation, Automatica 1995; 31, 1443-1451.
[4] Chen H, Allgower F. A quasi-innite horizon nonlinear predictive con-
trol scheme with guaranteed stability, Automatica 1998; 34, 1205-1217.
16
[5] De Nicolao G., Magni L., Scattolini R. Stabilizing receding-horizon
control of nonlinear time-varying systems, IEEE Transactions on Au-
tomatic Control 1998; 43, 1030-1036.
[6] Mayne DQ, Rawlings JB, Rao CV, Scokaert POM. Constrained model
predictive control: stability and optimality, Automatica 2000; 36, 789-
814.
[7] Magni L., De Nicolao G., Magnani L., Scattolini R. A stabilizing model-
based predictive control algorithm for nonlinear systems, Automatica
2001; 37, 1351-1362.
[8] Liu RW. Convergent systems, IEEE Transactions Automatic Control
1968; 13, 384-391.
[9] Angeli D, Casavola A, Mosca E. Constrained predictive control of non-
linear plants via polytopic linear system embedding, International
Journal of Robust and Nonlinear Control 2000; 10, 1091-1103.
[10] Sznaier M. Receding horizon: an easy way to improve performance in
LPV systems, Proc. American Control Conference 1999; San Diego,
USA, 4, 2262-2266.
[11] Chisci L, Falugi P, Zappa G. Predictive control for constrained LPV
systems, Proc. European Control Conference 2001; Porto, Portugal,
3074-3079.
[12] Casavola A, Famularo D, Franze G. A scheduling min-max predictive
control algorithm for LPV systems subject to bounded rate parameters.
Proc. 40th IEEE Conference on Decision and Control 2001; Orlando,
USA, 2372-2377.
[13] Shamma JS, Athans M. Gain scheduling: potential hazards and reme-
dies, IEEE Control Systems Magazine 1992; 12, 101-107.
[14] Rugh WJ, Shamma JS. Research on gain scheduling, Automatica 2000;
36, 1401-1426.
[15] Shamma JS, Xiong D. Set valued methods for linear parameter varying
systems, Automatica 1999; 35, 1081-1089.
[16] Rossiter JA, Kouvaritakis B, Rice MJ. A numerically robust state-space
approach to stable predictive control strategies, Automatica 1998; 34:
65-73.
17
[17] Artstein Z. Stabilization with relaxed control, Nonlinear analysis 1983,
7, 1163-1173.
[18] Lin Y, Sontag ED. A universal formula for stabilization with bounded
controls, Systems and Control Letters 1991, 16, 393-397.
[19] Freeman RA, Kokotovic PV. Inverse optimality in robust stabilization,
SIAM Journal Control and Optimization 1996, 34, 1365-1391.
[20] Freeman RA, Primbs JA. Control Lyapunov functions: new ideas from
an old source, Proc. 35th Conference on Decision and Control 1996;
Kobe, Japan, 3926-3931.
[21] Astrom KJ, Furuta K. Swinging up a pendulum by energy control,
Automatica 1999; 36, 287-295.
[22] Isidori A. Nonlinear Control Systems 3rd Edition 1995, Springer, Lon-
don, UK.
[23] Boyd S, El Ghaoui L, Feron E, Balakrishnan V. Linear Matrix Inequal-
ities in System and Control Theory 1994, SIAM, Philadelphia, USA.
[24] Barmisch RR. Necessary and sucient conditions for quadratic stabi-
lizability of an uncertain system, Journal of Optimization Theory and
Applications 1985; 46, 399-408.
[25] Geromel JC, Peres PLD, Bernussou J. On a convex parameter space
method for linear control design of uncertain systems, SIAM Journal
Control and Optimization 1991; 29, 381-402.
[26] Gilbert EG and Tan KT. Linear systems with state and control con-
straints: the theory and application of maximal output admissible sets,
IEEE Transactions Automatic Control 1991; 36, 1008-1021.
[27] Kolmanovski IV, Gilbert EG. Theory and computation of disturbance
invariant sets for discrete-time linear systems. Mathematical Problems
in Engineering: Theory, Methods, Applications 1998; 4: 317-367.
[28] Blanchini F. Ultimate boundedness control for uncertain discrete-time
systems via set-induced Lyapunov function, IEEE Transactions Auto-
matic Control 1994; 39, 428-433.
[29] Blanchini F. Set invariance in control. Automatica 1999; 35: 1747-
1767.
18
[30] Rao C, Rawlings JB, Wright S. Application of interior point methods
to model predictive control. Journal of Optimization Theory and Ap-
plications 1998: 99, 723-757.
Proof of Theorem 2 - Given S I IR
n
, let us denote by (t, S)
IR
n
the set of states x(t) originated from s(0) S for the LPV system
(18). If O
i
k
, i = 1, , , are bounded there exists a ball B IR
n
, contain-
ing the origin, such that O
i
k
B. By the exponential stability assumption
(t, I B) {0} as t . Hence there exists an index h > k for which
x (h + 1, I B) = (L
i
+ M
i
F
i
)x 1 for i = 1, 2, . . . , . Hence
O
h
O
k
I B and, therefore, x (h + 1, O
h
) = (L
i
+ M
i
F
i
)x
1 for i = 1, 2, . . . , . This, by denition implies that O
h
= O
h+1
and hence
O

= O
h
is nitely determined.
Proof of Theorem 3 - The hypotheses that x(t)

N
implies that
x(t)
i
N
for i(t) = i(x
1
(t)), i.e. s(t) = [i(t), x

(t)]


N
. Therefore
there exist control sequences c(t) such that [s

(t), c

(t)]

S
N
. Therefore
x(t)

N
implies that s(t + 1)
N1

N
. Hence, by induction,
s(t)
N
for all t 0 and satisfaction of the constraints (15) is guaranteed.
Next, consider the cost function V
t
= V (s(t))

= c(t)
2
. Since the
sequence {c(t +1|t), c(t +2|t), . . . , c(t +N 1|t), 0} is feasible at time t +1,
V
t
V
t+1
c(t)
2
0 (27)
where c(t)

= c(t|t). Hence {V
t
}
t0
is a nonnegative monotonic non-increasing
scalar sequence and, as t , must converge to V

< . Summing the


V
t
V
t+1
of (27), for t from 0 to , we have > V
0
V

t=0
c(t)
2

0 lim
t
c(t)
2
= 0 which proves that lim
t
c(t) = 0.
Any state x(t) reachable from x(0) under the non linear feedback (23) can
be expressed by
x(t) = (t, 0)x(0) +
t1

k=0
(t, k + 1)B(k)c(k) (28)
where (t, k)

= (t1) (k+1)(k) and (k) = A(k)+B(k)F
i(k)
, [A(k), B(k)]
F(i(k)) for any admissible sequence i(0), i(1), , i(k1). The contractivity
of
0
implies that there exists a constant M such that
|(t, k)| M
tk
(29)
19
From (28) and (29), it follows that lim
t
c(t) = 0 implies that lim
t
x(t) = 0.
20

Vous aimerez peut-être aussi