60 Asian Journal of Control, Vol. 12, No. 1, pp. 58 70, January 2010
such that:
L
l=1
l
(k) = 1, [A(k)|B(k)]=
L
l=1
l
(k)[A
l
|B
l
]. (3)
Our aim is to design MPC that brings the system
(1)–(2) to the origin (x = 0, u = 0) and at each time k
solves the following optimization problem:
min
u(k)
max
[A(k+i)|B(k+i)]∈,i≥0
J
∞
(k)
=
∞
i=0
[x(k + i|k)
2
Q
+u(k + i|k)
2
R
], (4)
s.t. x(k + i + 1|k)
=A(k + i )x(k + i|k) + B(k + i )u(k + i |k),
x(k|k) = x(k), i ≥ 0, (5)
−u
≤ u(k + i|k) ≤¯u,
−
≤ x(k + i + 1|k) ≤
¯
, i ≥ 0, (6)
where Q>0andR>0 are weighting matrices and
u(k) ={u(k|k), u(k + 1|k), u(k + 2|k), . . .} are the
decision variables. Problems (4)–(6) together are a
min-max optimization problem corresponding to a
worst-case infinite-horizon MPC.
The traditional open-loop min-max MPC param-
eterizes u(k) as N single control moves U (k) [20–23]
followed, in the simplest form, by
u(k + i|k) = F(k)x(k + i |k), i ≥ N
. (7)
In order to guarantee recursive feasibility, however, it
is advisable to utilize the single control moves only at
the initial time k = 0. For k>0, it is advisable to grad-
ually advance the single control moves into the partial
feedback form, as in [7].
The traditional open-loop min-max MPC may ex-
hibit excessive conservativeness due to the open-loop
nature of the prediction. Mayne et al. ([1], Section 4.6)
have suggested feedback MPC to improve feasibility
and optimality of the the traditional open-loop min-max
MPC, by parameterizing
u(k+i |k) = K (k+i|k)x(k+i|k), i = 1...N−1 (8)
for optimization. Define (k) := {u(k|k), K (k +
1|k),...,K (k + N − 1|k), F(k)}. Consider on-line
feedback MPC which, at each time k, solves the fol-
lowing optimization problem:
min
(k)
max
[A(k+i)|B(k+i)]∈,i≥0
J
∞
(k), s.t. (5)–(8). (9)
For N = 0, (k) := {F(k)}, Problem (9) has been solved
(e.g. [4]). By assuming the current [A(k)|B(k)] always
exactly known, [5] have further solved Problem (9)
with N = 1, (k) := {u(k|k), F(k)}. For general N ≥ 2,
the problem of calculating (k) has been regarded as
difficult.
III. UNIQUE FEATURE OF PDOLMPC
PDOLMPC avoids the difficulty of calculating
(k). Consider the state transfer with different pos-
sible vertex models, i.e. the models described by
A
l
, B
l
, l = 1 ...L. Define the vertex control moves
u(k|k), u
l
0
(k + 1|k),...,u
l
N −2
···l
0
(k + N − 1|k),
l
j
= 1 ...L, j = 0 ...N − 2, which satisfy:
x
l
0
(k + 1|k) = A
l
0
x(k) + B
l
0
u(k|k),
x
l
i
···l
0
(k + i + 1|k)
= A
l
i
x
l
i−1
···l
0
(k + i|k) + B
l
i
u
l
i−1
···l
0
(k + i|k),
i = 1 ...N − 1, l
j
= 1 ...L,
j = 0 ...N − 1,
where x
l
i−1
···l
0
(k + i|k), i = 1 ...N are vertex state pre-
dictions.
The real control move u(k + i|k) for any i>0is
uncertain and defined by:
u(k + i|k) =
L
l
0
···l
i−1
=1
i−1
h=0
l
h
(k+h)
×u
l
i−1
···l
0
(k+i|k)
,
L
l
0
···l
i−1
=1
i−1
h=0
l
h
(k + h)
= 1,
i = 1 ...N − 1. (10)
According to (10), u(k +i |k), i = 1 ...N −1 are param-
eter dependent, i.e. each u(k +i|k) is a convex combina-
tion of all the items u
l
i−1
···l
0
(k + i|k) with the unknown
combining coefficients
i−1
h=0
l
h
(k + h), l
j
= 1 ...L,
j = 0 ...i −1. PDOLMPC also avoids calculating u(k+
i|k), i>0.
Since [A(k)|B(k)] belongs to a convex hull, the
model prediction (5) can be expressed as
x(k + 1|k) = A(k)x( k|k) + B(k)u(k|k)
=
L
l
0
=1
l
0
(k)[A
l
0
x(k) + B
l
0
u(k|k)]
=
L
l
0
=1
l
0
(k)x
l
0
(k + 1|k)
q
2009 John Wiley and Sons Asia Pte Ltd and Chinese Automatic Control Society