Dynamic output feedback robust MPC for constrained
quasi-LPV systems with both polyhedral and ellipsoidal bounds
on estimation error
DING Baocang
Ministry of Education Key Lab For Intelligent Networks and Network Security (MOE KLINNS Lab), Department of Automation, School
of Electronic and Information Engineering, Xi’an Jiaotong University, Xi’an 710049, P. R. China
E-mail: baocang.ding@gmail.com
Abstract: A new approach is proposed for the output feedback robust model predictive control (MPC) of a quasi-linear parameter
varying (quasi-LPV) system with bounded noise. In the main optimization problem for finding the control law, the current true
state is substituted with its bounds. The bounds are represented by either an ellipsoid or a polyhedron, the latter being used if and
only if it is contained in the former. Both the ellipsoid and the polyhedron are refreshed at each sampling time. The recursive
feasibility of the main optimization problem is guaranteed by a simple refreshment of the ellipsoid, the control performance can
be improved by applying the polyhedron, and the closed-loop stability is achieved by the technique of quadratic boundedness. A
numerical example is given to illustrate the effectiveness of the controller.
Key Words: Dynamic output feedback, Model predictive control, Uncertain systems, Quadratic boundedness
1 INTRODUCTION
The output feedback robust model predictive control
(OFRMPC) for the discrete-time linear system has been ex-
tensively studied. When the bounded noise exists but the
model parametric uncertainty does not, OFRMPC has been
successful (e.g. [10, 12–14]). When both the bounded noise
and the model parametric uncertainty exist, OFRMPC has
been tried (e.g. [4,6–9]). While [6–9] consider the linear pa-
rameter varying (LPV) system, [4] considers the quasi-LPV
system. Differently from a usual LPV system, in a quasi-
LPV system the varying parameters of the system are exactly
known at the current time. Reference [5] studies the Takagi–
Sugeno fuzzy system which, in this context, is similar to the
quasi-LPV system.
A key issue for OFRMPC is the computation of the outer
bounds of true state. In [4–9], these bounds are calculated
by adding the estimated state to the estimation error bounds
(EEBs). In [8] and [9], EEBs are fixed, by imposing an esti-
mation error constraint in the optimization problem. Adding
this artificial constraint brings conservatism with respect to
feasibility and performance. Since there is no refreshmen-
t of EEBs, conservatism is accumulated with the evolution
of time. In [4–7], there is no estimation error constraint in
the main optimization problem for finding the control law,
and EEBs are refreshed at each sampling time perhaps by an
auxiliary optimization problem.
In [8] and [9], off-line model predictive control (MPC),
where all the optimizations are performed off-line, is uti-
lized. For off-line MPC, there is no requirement for the re-
cursive feasibility: a property showing that MPC optimiza-
tion problem is feasible for all future time when it is feasi-
ble at the initial time [15]. In [4], at each sampling time,
the main optimization problem may or may not be solved,
depending on the previous auxiliary optimization problem.
This work was supported by the National Nature Science Foundation
of China under Grant 60934007 and Grant 61174095, and by the Founda-
tion from the State Key Laboratory of Industrial Control Technology under
Grant no. ICT1213.
This can be seen as a new method for avoidance of the re-
quirement for recursive feasibility. In [5–7], a simple method
is proposed for guaranteeing the recursive feasibility.
This paper contributes on the refreshment and usage of
EEBs, in the case of guaranteed stability. In [4], the poly-
hedral set of the estimation error is refreshed by a positive-
definite matrix. In this paper, EEBs are refreshed by consid-
ering both an ellipsoid and a polyhedron, the latter being less
conservative than that in [4]. The ellipsoid introduces the cri-
terion for recursive feasibility, as in [5–7]. Differently from
the existing works, both the ellipsoid and the polyhedron are
used in the main optimization problem.
In the linear matrix inequality (LMI) optimization prob-
lems of [8] and [9], some controller parameters are pre-
specified, in order to reduce the computational burden. In [6]
and [7], there is no pre-specification of any controller param-
eter, but the solution procedure for the optimization prob-
lem is computationally very expensive. In comparison, [4]
and [5] do not pre-specify any controller parameter, and the
involved computation is considerably lighter than that in [6]
and [7]. This paper adopts the system the same as in [4].
The new idea in this paper can be easily applied to the sys-
tem and technique in [5–7], having no contribution to the
computational issue. The differences between this paper and
the existing works are shown in Table 1.1.
Notations: I is the identity matrix with appropriate di-
mension. For time k, we often use := k +1for simplicity.
The time-dependence (k) of the MPC decision variables is
often dropped for simplicity. For any vector x and positive-
definite matrix W, x
2
W
:= x
T
Wx, ε
W
:= {ξ|ξ
T
Wξ ≤
1}, and x(i|k) is the value of x at time k + i predicted at
time k. For a set S and a real matrix M of compatible di-
mension, MS := {Mx|x ∈S}. All vector inequalities are
interpreted in an element-wise sense. An element belonging
to CoS means that it is a convex combination of the elements
in S, with the scalar combining coefficients nonnegative and
their sum equal to 1. A value with superscript ∗ means that
it is the optimal solution of the optimization problem. An
Proceedings of the 32nd Chinese Control Conference
Jul
26-28, 2013, Xi'an, China
4143