Beruflich Dokumente
Kultur Dokumente
S. OLARU, D. DUMUR
Abstract: Model predictive control (MPC) has the ability to explicitly handle inequality
constraints on process variables due to its time-domain formulation. However, the
controller is not available off-line, leading to a nonlinear control law. This paper
examines the stability in the MPC framework viewed as consequences of the feasibility
of the constrained predictive law. If the current results in the literature deal with on-line
feasibility conditions, this paper is providing tuning rules and performance analysis
based on off-line prediction of infeasibility.
−1 −1
Aγ (q ) γ (t ) = Bγ (q ) u (t − d )
γ min ≤ γ (t + j t ) ≤ γ max , N c1 ≤ j ≤ N c 2
(6)
w
Fig. 1. Constrained predictive controller principle
Based on the model mentioned earlier and following
the ideas of GPC (Generalized Predictive Control) Controller
N c1 and N c 2 represent the constraints horizon
and can be different from the prediction horizon (
N1, N 2 ) or postulated control horizon N u . This
+
(Clarke, et al., 1994) an optimal j-step ahead predict-
tor can be constructed:
description form allows the consideration of all types
of constraints rising from operational restriction or
−1
−1 -
imposed by stability considerations, e.g. equality
yˆ (t + j ) = F j ( q ) y (t ) + H j ( q ) ∆u (t −1) + constraints on the output for m instants at the end of
the prediction horizon, inequality constraints on the
free response output, inequality constraints on the control action,
+ G j ( q −1 ) ∆u (t + j −1) inequality constraints on the control increments.
Model based
forced response
(2) 3. STABILITY AND FEASIBILITY
where the F j , G j , H j polynomials are solutions The constraint predictive law minimizes a cost
function in the control vector ∆u defined on a
of the Diophantine equations:
constrained
∆(q −1 ) A(q −1 ) J j (q −1 ) + q − j F j (q −1 ) = 1
domain described by a linear inequalities system. The
domain for the solution of this quadratic program can
be either a nonempty or empty convex set. The
predictive
−1 −j −1 −1 −1
G j ( q ) + q H j ( q ) = B ( q ) J j (q )
assumption of feasibility is equivalent with the fact
Predic
(3) that this domain is non empty. It was shown that a
close relationship exists between feasibility and
controller
The closed loop stable behaviour is achieved by
applying the first control action of the sequence stability of closed loops with constrained predictive
which minimizes a quadratic cost function: control (The stability is related to BIBO definition;
the possibility of non damped harmonics remains and
more refined stability results may be applied). If
N2 Nu studies start with the feasibility assumption, results
J = ∑ [ w(t + j ) − yˆ (t + j )] 2 + λ [ ∆u (t + j −1)concerning
∑ ]2 stability are somehow direct when infini-
j =N1 j =1 te or pseudo-infinite constraint horizons are used.
(4) Note that stability as a consequence of feasibility
assumes the absence of disturbances. In case of
subject to ∆u (t + j ) = 0 for j ≥ N u . N1 , N 2
disturbances, they must not alienate feasibility.
are the minimum and maximum costing horizon,
Consider an open loop stable system with upper and compact sets and a pseudo-infinite upper constraint is
lower bounds on any constrained variable γ(t ) used, the infeasibility is preceding the instability.
corresponding to a minimum phase system of Eq. 6.
This argument clears the demarcation between two
For any setpoint keeping the feasibility assumption
distinct kinds of problems: feasible and stable against
alive, the quantities u , ∆u or y remain in a close
unfeasible and unstable. If the set of constraints is
set implying BIBO stability (Scokaert and Clarke,
chosen in the sense of a pseudo-infinite prediction
1994a). Some difficulties may appear when the
horizon, the frontiers of feasible and stable control
constraints are stated to a non-minimum phase
laws is strict. In counterpart even if infeasibility does
variable γ(t ) . This is because constraining this
not mean directly instability, it may be considered as
variable in its transient period leads to instability as
harmful. In this case, special strategies must be taken
the feasibility demands an “inverse” control action
in consideration in order to continue control and this
and on a long-term creates a snowball effect driving
procedure is to be avoided. So, when dealing with
the control action and the entire system unstable. The
feasibility, one has to inspect first the optimization
problem is important when the constrained variable is
domain represented by a polyhedron in the N u
disturbed, and can be overcome by modifying the
dimensional space. In the case of pseudo-infinite
constraint horizons. From the precedence point of
constraint horizons, this polyhedron is defined by the
view, when the constraint horizon is long enough, the
end-point equalities and the inequality constraints
infeasibility will precede the instability. As an
and should be non-empty to consider the quadratic
example, consider a non-minimum phase system
problem as feasible. It is however possible that some
under the constraint −0.3 ≤ y ≤1 and a step
of the constraints conflict and become incompatible.
setpoint:
The following sections will treat extensively the
(1 − q −1 + 0.25 q −2 )γ (t ) = feasibility problem. Distinguish at the beginning, as
( −0.25 − 0.25 q −1 + 0.75 q −2 )u (t ) in (Scokaert and Clarke, 1994b):
(7) – Type I infeasibility: caused by an incompa-
tibility between inequality constraints,
For a constraint prediction horizon of a single step
iteration N c1 = N c 2 = 1 , the closed loop system is – Type II infeasibility: caused by an incompa-
tibility between the end-point equality, the
feasible but unstable as shown in Figure 2.
system dynamics and inequality constraints.
More difficult is the type II infeasibility because it
arises when the end-point equalities are employed for
stability reasons and conflict with the system
dynamics and the other inequality constraints. A first
consequence is that the setpoint may interfere with
the constraints definitions. Even if this new problem
Fig. 2. Unstable behaviour for feasible control law. can not be treated off-line extensively we can
Increasing the constraint prediction horizon, however made some considerations. Notice that for a
N c 2 = 6 , infeasibility is signalled at the 6th given set of inequality constraints, a step setpoint can
iteration confirming the fact that if the horizon is always be found such that equality constraint at the
long enough infeasibility is preceding instability. If end of the prediction horizon generates infeasibility.
by off-line analysis the infeasibility is predicted, then Theorem 1: A CRHPC control law with inequality
rules for choosing the constraint horizon can be constraints for a system with a step setpoint starting
inferred to achieve feasibility. Figure 3 presents the at time t is feasible at each instant t +k if:
simulations for N c1 = 1 and N c 2 = 12 . Further, i) The problem is feasible at instant
feasibility implies the stability of the control law. □ t;
ii)
m ≥ n a = deg( A) −1; N u ≤ N 2 +1 + d
Proof: Consider the optimal control sequence at
instant t:
k u = { ∆u (t + j −1 / t )}, j = 1, , N u , obtained
by minimization of the cost function Eq. 4 under the
Fig. 3. Feasibility and stability. constraints stated in Eq. 6. If the problem is feasible
at instant t , the optimal control sequence k u is
With unstable open loop systems, the problem is available. Consider the suboptimal control vector:
much more complicated but there are motivations to
think that if all manipulated signals are restricted to k u ∗ = { ∆u (t + j − 1 / t )} , j = 2, , N u (8)
Without disturbances, all predictions of the constraint missing the generality and have to be avoided. The
variables satisfy: output y, even if linked to the input sequence k u ,
γ (t + j / t +1) = γ (t + j / t ) (9) will thus be present in the matrix form of the
constraint as an unknown variable. Its freedom is
This mean that all the constraints are satisfied for limited by the dynamic constraints. This means that
k u ∗ . In the same manner if the control actions the new variables vector is:
postulated at time t +1 is k u ∗ , the output
y (t + j / t +1) = y (t + j / t ) .
1 − 1 0 0 (ty + N2 + 1) 0
equations and inequalities
{x Aeq x = Beq ; Ain x ≤bin }defining a
polyhedron P which can further be expressed by its
dual representation:
0 1 − 1 0 (ty + N + 2) 0
P = conv.hull{ x1, , xv } +
+ cone{ y1, , yr } + lin.spaceP
(18)
2 =
where conv.hullX denotes the set of all convex
combinations of points in X, coneY denotes
nonnegative combinations of unidirectional rays and
lin.spaceP represents a linear combination of bi-
directional rays. It can be rewritten as:
0
v r l
P = ∑λi x i + ∑γ i y i + ∑µi z i
(16) i =1 i =1 i =1
v (19)
0 ≤ λi ≤ 1, ∑λi = 1 , γ i ≥ 0 , ∀µi
i =1
4q −1 + 6.743 q −2
H ( q −1 ) =
1 − 0.82 q −1 − 0.16 q −2 − 0.017 q −3
J yc −1 ≤ ∆u i ≤ 0.1 1 ≤ i ≤ N u
−1 ≤ u i ≤1 1 ≤ i ≤ N u
−5 ≤ y i ≤ 5 N1 ≤ i ≤ N 2
(20)
With these particular descriptions, a feasibility
All this requires the constrained variables set
domain for a system under all types of constraints is: θ = [∆u1 ∆u Nu y N1 y N 2 ] . With N u = 3
G − I 0 ∆u − l , N 1 = 1 , N 2 = 6 , the dual representation of this
is a six points convex hull given in Figure 4.
G c 0 − I y = − l c θ1=[32 32 -139 128 576 628 -186
0 0 J y c 0 -893 -1600]/320
θ2=[32 32 -39 128 576 1028 1214
Aeq θ Beq 1407 1600]/320
∆u I 0 0 ∆u (17) θ3=[-22 49 49 -88 -112 376 1109
y 0 I 0 ∆u y 1815 2450]/490
≤ y ≤ θ4=[40 -31 40 160 436 646 1127
0 0 I y c
1528 2000]/400
yc
y c θ5=[-122 49 49 -488 -1512 -1924 -2091
-2185 -2450]/490
u Q 0
θ0
u
θ6=[40 -131 40 160 36 -754 -1173
bi Ai bs -1672 -2000]/400
[ ]'
with y = y N1 , , y N 2 , x and x the upper
Fig. 4. Convex hull computed by POLYLIB
and lower bounds of the respective variables. When The same system under some unfortunate constraints
the domain is totally determined a specific setpoint like 2 ≤ y i ≤ 5 N1 ≤ i ≤ N 2 will return an empty
will be compared with the reference trajectory domain which means infeasibility of the constraint
bounds in order to check its feasibility. predictive problem whatever the setpoint is.
4.2.Type I infeasibility 4.3.Type II infeasibility for step input response
Type I infeasibility is easy to check by inspection of Section 4.2 has proposed a technique to determine a
the inequalities or system dynamics. A possible way type I feasible domain (a polyhedron equivalent to
the set of constraints). This convex set is providing a presented in section 4.2 provides the results of Figure
starting point for type II infeasibility analysis. It tries 5, displaying the setpoint magnitude as a function of
to find an upper (resp. lower) bound for feasible step m, N u and N 2 .
setpoints. Section 3 has shown that all feasible steps
at initial instant preserve this property for all further
time instants. Thus the bounds of the feasible step
setpoints are determined as a linear program in the
type I feasible domain. As noticed in (Le Verge,
1992), the algorithm finding the dual representation
of a polyhedron is used to find the set of all solutions
Fig. 5. Feasible step setpoints.
to a linear programming problem, in our case:
The evolution of the upper bounds for the feasible
step setpoints is presenting important differences in
max [0 0 1] θ Aeq θ =B eq ; bi ≤ Ai θ ≤b s the case m = 1 and m = 2 compared with
c
m > 2 . A closer look will reveal that whatever the
(21) choice of parameters N u and N 2 is, the
monotonicity of the feasible steps with respect to
Three cases may arise: 1) if the associated poly-
N 2 is lost as long as m < n a ( n a = 3 in the
hedron is empty, the problem is infeasible; 2) if there
example). This is an impor-tant aspect because in
exists a bi-directional ray z such that c z ≠ 0 or a
real-time applications, N u is kept as low as
unidirectional ray y such that c y > 0 , then the
possible and the real tuning parameter is N 2 .
maximum is unbounded; 3) if all bi-directional rays z
Adding to this the fact that, as shown in Theorem 1,
are such that c z =0 and all unidirectional rays y
for step setpoints the feasibility proper-ties are
are such that c y ≤ 0 , then the maximum is defined
inherited for all future instances if m ≥ n a , gives a
by max {c x x vertex of P} , and the set of
first result for the best control parameters.
solutions is:
{ } { }
G c is a ( m × N u ) matrix with rank strictly
P = conv .hull x11, , x1v + cone y11, , y1r + lin .related
spaceP with n a . which means that equality
1 constraints at the end of the prediction horizon are
where x1 are the vertices attaining this maximum
effective only for:
and y1 are such that cy 11 = 0 . This result is used
1