Introduction to Constrained Control

advertisement
Introduction to Constrained Control
Graham C. Goodwin
September 2004
Centre for Complex Dynamic
Systems and Control
1.1 Background
Most of the literature on Control Theory deals with Linear
Unconstrained Systems.
However to get the most out of a system, we usually need to
deal with nonlinearities.
The most common nonlinearity met in practice are Actuator
Limits.
Centre for Complex Dynamic
Systems and Control
To get the most out of a system you need to push up against limits.
Centre for Complex Dynamic
Systems and Control
Other examples?
Playing sport at international level
Excelling in business or academia
Aerospace, chemical process control, . . .
Control is a key enabling technology in many (all?) areas
Getting the most out of control means pushing against boundaries
Centre for Complex Dynamic
Systems and Control
1.2 Approaches to Constrained Control
Cautious
(back off performance demands so constraints are not met)
Serendipitous
(allow occasional constraint violation)
Evolutionary
(begin with a linear design and add embellishments, for
example, antiwindup)
Tactical
(include constraints from the beginning, for example, MPC)
Centre for Complex Dynamic
Systems and Control
1.3 Example: Rudder Roll Stabilisation of Ships (See
lecture 3.5)
It has been observed that unless appropriate actions are taken to
deal with constraints, then the performance of rudder roll
stabilisation systems can be worse than if nothing is done due to
the effect of actuator amplitude and slew rate constraints.
Centre for Complex Dynamic
Systems and Control
1.4 Model Predictive Control
Model Predictive Control (MPC)
is a prime example of tactical method.
Long history in petrochemical industry.
Many thousands of applications.
Several commercial products.
Industrial “credibility”.
Centre for Complex Dynamic
Systems and Control
Background
A survey by Mayne et al. (2000) divides the literature on MPC in
three categories:
Theoretical foundations: the optimal control literature.
Dynamic Programming (Bellman 1957), the Maximum
Principle (for example, Lee & Markus 1967).
“Process control” literature, responsible for MPC’s adoption by
industry. Evolving generations of MPC technology. An
example of practice leading theory.
“Modern” literature, dealing with theoretical advances such as
stability and robustness.
Centre for Complex Dynamic
Systems and Control
General Description
MPC is a control strategy which, for a model of the system,
optimises performance (measured through a cost function)
subject to constraints on the inputs, outputs and/or internal
states.
Due to the presence of constraints it is difficult, in general, to
obtain closed formulae that solve the above control problem.
Hence, MPC has traditionally solved the optimisation problem on
line over a finite horizon using the receding horizon technique.
This has also restricted the applicability of MPC to processes with
slow time constants that allow the optimisation to be solved on line.
Recent results allow faster systems to be handled.
Centre for Complex Dynamic
Systems and Control
An Illustrative Example
We will base our design on linear quadratic regulator [LQR] theory.
Thus, consider an objective function of the form:
VN ({xk }, {uk }) ,
N −1
1 
1 X 
xN PxN +
xk Qxk + uk Ruk ,
2
2
(1)
k =0
where {uk } denotes the control sequence {u0 , u1 , . . . , uN −1 }, and
{xk } denotes the corresponding state sequence {x0 , x1 , . . . , xN }.
In (1), {uk } and {xk } are related by the linear state equation:
xk +1 = Axk + Buk ,
k = 0, 1, . . . , N − 1,
where x0 , the initial state, is assumed to be known.
Centre for Complex Dynamic
Systems and Control
The following parameters allow one to influence performance:
the optimisation horizon N
the state weighting matrix Q
the control weighting matrix R
the terminal state weighting matrix P
For example, reducing R gives less weight on control effort, hence
faster response. R → 0 is called “cheap control”.
Centre for Complex Dynamic
Systems and Control
Details of Example
Consider the specific linear system:
xk +1 = Axk + Buk ,
(2)
yk = Cxk ,
with
"
1
A=
0
#
1
,
1
"
#
0.5
B=
,
1
h
C= 1
i
0 ,
which is the zero-order hold discretisation with sampling period 1
of the double integrator
d 2 y (t )
= u(t ).
dt 2
Centre for Complex Dynamic
Systems and Control
Example
eplacements
controller
uk
sat
xk
linear
system
Figure: Feedback control loop for Example



1
if u > 1,




sat(u) , 
u
if |u| ≤ 1,




−1 if u < −1.
(3)
Centre for Complex Dynamic
Systems and Control
(i) Cautious Design
"
1
(N = ∞, P = 0) and weighting matrices Q = C C =
0
R = 20 gives the linear state feedback law:

h
uk = −Kxk = − 0.1603
#
0
and
0
i
0.5662 xk .
Centre for Complex Dynamic
Systems and Control
Cautious Design
1
0.8
uk
0.6
0.4
0.2
0
−0.2
−0.4
0
5
10
k
15
20
25
0
5
10
k
15
20
25
1
0
−1
−2
yk
ag replacements
−3
−4
−5
−6
Figure: uk and yk for the cautious design uk = −Kxk with weights Q = C  C
and R = 20.
Centre for Complex Dynamic
Systems and Control
(ii) Serendipitous Design
Using the same Q = C  C in the infinite horizon objective function
we try to obtain a faster response by reducing the control weight to
R = 2. We expect that this will lead to a control law having “higher
gain.”
Centre for Complex Dynamic
Systems and Control
Serendipitous Design
3
uk
2
1
0
−1
0
5
10
k
15
20
25
0
5
10
k
15
20
25
1
0
−1
−2
yk
ag replacements
−3
−4
−5
−6
Figure: uk and yk for the unconstrained LQR design uk = −Kxk (dashed
line), and for the serendipitous strategy uk = −sat(Kxk ) (circle-solid line),
with weights Q = C  C and R = 2.
Centre for Complex Dynamic
Systems and Control
Encouraged by the above result, we might be tempted to “push our
luck” even further and aim for an even faster response by further
reducing the weighting on the input signal. Accordingly, we
decrease the control weighting in the LQR design even further, for
example, to R = 0.1.
Centre for Complex Dynamic
Systems and Control
6
4
uk
2
0
−2
−4
−6
0
5
10
k
15
20
25
0
5
10
k
15
20
25
4
2
yk
ag replacements
0
−2
−4
−6
Figure: uk and yk for the unconstrained LQR design uk = −Kxk (dashed
line), and for the serendipitous strategy uk = −sat(Kxk ) (circle-solid line),
with weights Q = C  C and R = 0.1.
Centre for Complex Dynamic
Systems and Control
The control law u = −sat(Kx ) partitions the state space into three
regions in accordance with the definition of the saturation
function (3). Hence, the serendipitous strategy can be
characterised as a switched control strategy in the following way:



−Kx




u = K(x ) = 
1




−1
if x ∈ R0 ,
(4)
if x ∈ R1 ,
if x ∈ R2 .
Notice that this is simply an alternative way of describing the
serendipitous strategy since for x ∈ R0 the input actually lies
between the saturation limits. The partition is shown in following
figure.
Centre for Complex Dynamic
Systems and Control
Figure 5
4
3
2
xk2
1
0
ag replacements
−1
R2
−2
R0
−3
−4
−6
R1
−4
−2
xk1
0
2
4
6
Figure: State space trajectory and space partition for the serendipitous
strategy uk = −sat(Kxk ), with weights Q = C  C and R = 0.1.
Centre for Complex Dynamic
Systems and Control
Examination of figure 8 suggests a heuristic argument as to why
the serendipitous control law may not be performing well in this
case. We can think, in this example, of x 2 as “velocity” and x 1 as
“position.” Now, in our attempt to change the position rapidly (from
−6 to 0), the velocity has been allowed to grow to a relatively high
level (+3). This would be fine if the braking action were
unconstrained. However, our input (including braking) is limited to
the range [−1, 1]. Hence, the available braking is inadequate to
“pull the system up”, and overshoot occurs.
Centre for Complex Dynamic
Systems and Control
(iii) Tactical Design
Perhaps the above heuristic argument gives us some insight into
how we could remedy the problem. A sensible idea would seem to
be to try to “look ahead” and take account of future input
constraints (that is, the limited braking authority available). To test
this idea, we take the objective function (1) as a starting point.
Centre for Complex Dynamic
Systems and Control
Tactical Design
We use a prediction horizon N = 2 and minimise, at each sampling
instant i and for the current state xi , the two-step objective function:
i +1
1 X 
1
V2 ({xk }, {uk }) = xi+2 Pxi +2 +
xk Qxk + uk Ruk ,
2
2
(5)
k =i
subject to the equality and inequality constraints:
xk +1 = Axk + Buk ,
|uk | ≤ 1,
(6)
for k = i and k = i + 1.
Centre for Complex Dynamic
Systems and Control
Tactical Design
In the objective function (5), we set, as before, Q = C  C , R = 0.1.
The terminal state weighting matrix P is taken to be the solution of
the Riccati equation P = A  PA + Q − K  (R + B  PB )K , where
K = (R + B  PB )−1 B  PA is the corresponding gain.
Centre for Complex Dynamic
Systems and Control
Tactical Design
As a result of minimising (5) subject to (6), we obtain an optimal
fixed-horizon control sequence {ui , ui +1 }. We then apply the
resulting value of ui to the system. The state evolves to xi +1 . We
now shift the time instant from i to i + 1 and repeat this procedure.
This is called receding horizon control [RHC] or model predictive
control.
Centre for Complex Dynamic
Systems and Control
Receding Horizon Technique
(1) At time i and for the current state xi solve an
open-loop (OL) optimal control problem over a
prediction horizon using a model of the system
to predict future states and taking into account
the present and future constraints;
(2) Apply the first step of the resulting optimal OL
control sequence;
0 1
2
3
k
0 1
2
3
k
(3) “Move the horizon”, that is, repeat the procedure
at time i + 1 for the current state xi + 1.
0 1
2
3
4
k
Centre for Complex Dynamic
Systems and Control
6
4
uk
2
0
−2
−4
−6
0
5
10
k
15
20
25
0
5
10
k
15
20
25
4
2
yk
ag replacements
0
−2
−4
−6
Figure: uk and yk for the unconstrained LQR design uk = −Kxk (dashed
line), and for the receding horizon design (circle-solid line), with weights
Q = C  C and R = 0.1.
Centre for Complex Dynamic
Systems and Control
We will see later that the receding horizon strategy described
above also leads to a partition of the state space into different
regions in which affine control laws hold. The result is shown (for
interest) in figure 9. The region R2 corresponds to the region R2 in
figure 8 and represents the area of state space where u = −1 is
applied. Comparing figure 8 and figure 9 we see that the region R2
has been “bent over” in the figure 8 so that u = −1 occurs at lower
values of x 2 (velocity) than was the case in figure 8. This is in
accordance with our heuristic argument about “needing to brake
earlier.”
Centre for Complex Dynamic
Systems and Control
Figure 7
4
3
R3
R2
2
R0
1
xk2
g replacements
0
−1
R1
−2
R4
−3
−4
−6
−4
−2
xk1
0
2
4
6
Figure: State space plot for the receding horizon tactical design.
Centre for Complex Dynamic
Systems and Control
Figure 5 and Figure 7
4
4
PSfrag replacements
3
2
1
1
R3
R2
R0
xk2
2
xk2
ents
3
0
0
−1
−1
R2
R1
−2
−2
R4
R0
−3
−4
−6
−3
R1
−4
−2
0
xk1
2
4
6
Figure: State space trajectory and
space partition for the serendipitous
strategy uk = −sat(Kxk ), with weights
Q = C  C and R = 0.1.
−4
−6
−4
−2
0
xk1
2
4
6
Figure: State space plot for the
receding horizon tactical design.
Centre for Complex Dynamic
Systems and Control
Summary
Can often avoid constraints by lowering performance
demands
However, this is at a cost
If we increase demands - constraints are met
Small violations not too significant
Soon get poor performance
Rethink the problem - add constraints into the design
Leads to idea of Receding Horizon Control
Centre for Complex Dynamic
Systems and Control
Download