Professional Documents
Culture Documents
net/publication/228671187
CITATIONS READS
6 2,588
3 authors, including:
All content following this page was uploaded by Gregor Gregorcic on 10 June 2014.
II Adaptive Control Vs. Local Model This control structure can be represented as shown
Network Control in Fig. 1.
following ARX model of the process [4]. Here u(k) Set−Point Controller
Feedforward Nonlinear System
and y(k) are the input and output of the process, Controller
Feedback
S(q)
R(q) y(k)
with A(q) and B(q) polynomials in the forward u(k)
RLS theta(k+1)
shift operator, ε(k) is white noise. Recursive Least Squares Estimator Controller Update
A(q) = q n + an−1 q n−1 + an−2 q n−2 + · · · + a0 (2) The closed–loop system can be represented as fol-
lows:
(A(q)R(q) + B(q)S(q)) y(k) = B(q)T (q)r(q) (8)
B(q) = Bm q m + bm−1 q m−1 +
Pole–Placement design is the choice of polynomi-
bm−2 q m−2 + · · · + b0 (3)
als R(q) and S(q), for estimated Â(q) and B̂(q)
polynomials, to obtain some desired closed–loop
This model can be continuously identified using,
characteristic equation Acl (q). This is represented
for example, the Recursive Least Squares algo-
by the following Diophantine equation:
rithm, with exponential weighting [4]:
Acl (q) = Â(q)R(q) + B̂(q)S(q) (9)
θ̂(k + 1) = θ̂(k) + P(k + 1)ψ(k + 1)
h i If polynomials Â(q) and B̂(q) do not have com-
y(k + 1) − ψ T (k + 1)θ̂(k) (4) mon factors, then this Diophantine equation can
be solved using matrix calculations, based on the
Sylvester matrix [2].
P(k + 1) = b) Multiple Model Network Control
" T
#
P(k)ψ(k + 1)ψ (k + 1)P(k) 1 When a single identification model is used, it will
= P(k) − (5)
λ + ψ T (k + 1)P(k)ψ(k + 1) λ have to adapt itself to the operating condition be-
fore appropriate control can be taken. After the
where the estimate of the parameter vector is de- operating condition of a system changes abruptly,
fined as: the original model (and hence controller) is no
longer valid. Slowness of adaptation may result
θ̂(k) = [ân−1 (k), · · · , â0 (k), b̂m (k), · · · , b̂0 (k)]T (6) in a large transient error.
If models are available for different operating
Exponential weighting implies that old measure- conditions, controllers corresponding to them can
ments are exponentially discounted, which gives be found in advance. The control strategy shown
an ability to track changes in time–varying and in [9], [10] is to determine the best model for the
nonlinear systems. The parameter λ in equation current operating condition at every instant, and
(5) is called the forgetting factor and is usually activate the corresponding controller. This control
chosen in range 0.95 < λ < 1. The choice of λ de- structure, as shown in Fig. 2, is based on N mod-
pends on how the properties of the system change. els which have been determined at various points
Smaller values of λ result in a faster forgetting, across the operating range of the process. A con-
which can be used to cope with nonlinear and fast troller is then designed for each model, using the
changing systems. Likewise values of λ close to Diophantine pole–placement algorithm. A super-
1 result in slower forgetting, and can be used for visor then compares the output errors for the N
systems that change gradually. models. A discrete equivalent of the performance
The control law of equation (7) is assumed, index proposed in [10] is given in equation (10), for
where R(q), S(q) and T (q) are polynomials in the the ith model:
forward shift operator. u(k) and y(k) are the plant M
X
input and output, r(k) is the set–point. To have Ji (k) = αe2i (k) + β exp(−jλi )e2i (k − j) (10)
a causal controller, it is necessary that the degree j=1
However, for a nonlinear process operating over
its full range, the steady–state control input at
each operating point is usually a nonlinear func-
tion of the desired output. It is therefore difficult
to provide zero mean data without online estima-
tion of the means. It is then preferable to use the
following second order discrete time incremental
model [12]:
Here ei (k) is the output error for the ith model. ∆Ca (k) = 1 − q −1 Ca (k)
(15)
The parameters α, β and λ are chosen to yield
a desired combination of instantaneous and long-
term accuracy. The forgetting factor λ determines
∆qc (k) = 1 − q −1 qc (k)
(16)
the memory of the index, and will have direct effect
on the decision speed of the supervisor. The model This will provide the additional benefit of intro-
with the lowest cost is selected, and its associated ducing integral action into a resulting controller.
controller is switched in the loop. For good parameter estimation, the system should
III The Nonlinear System be excited with a signal of rich frequency content.
PRBS PRBSsw
1 0.00019z−1+0.00016z−2
PRBS @ Input
1−1.59z−1+0.71z−2
V RT (t) 0.08
Constant
1+0.54z−1
+1.28e−007z−2
Filter
2
1
1−z−1
PLANT
k3 P Matrix: 0
Switch
Delta Operator
CSTR
1+0.54z−1
+1.28430871649e−007z−2 Control Panel Display
0.11
0.00018z−1
+0.00018z−2
1−1.87z−1
+0.94z−2
Model_0.13 0.08
0.00018z−1
+0.00017z−2
supervisor
1−1.69z−1
+0.78z−2 54 56 58 60 62 64 66 68 70
S−Function
Model_0.105
Time (min)
0.00018z−1
+0.00015z−2
1−1.48z−1
+0.59z−2
691.98
1+0.65z−1
+9.82e−008z−2 Model_0.08
qc_ini Constant1
Filter_0.13
1
Fig. 7: Set–point tracking for the CSTR using adaptive
710.10 PLANT
1
Constant2
0.08
Set Point
1+0.59z−1
+1.17e−007z−2
Filter_0.105 Multiport
1−z−1
Filter_0.08
Feedback_0.13 was defined as second order dominant with a damping
rad
5048.55−7088.15z−1+2749.70z−2
1+0.59z−1
+1.17e−007z−2
ratio of 0.707 and natural frequency of 7.4 min
Multiport Feedback_0.105
Switch2
4344.24−5629.5z−1
+2037.52z−2
1+0.52z−1
+1.54e−007z−2
Feedback_0.08
ing point, the parameters can be stored in to the
Multiple Model block diagram. This is done be
Fig. 6: Multiple Model block diagram switching the Multiple Model Tuning switch in the
Control Panel to on. It is preferable, that adapta-
The Supervisor block calculates the cost, proposed tion (RLS switch) is turned off, while the model
in (10), at each sampling time. The model with the and controller transfer function in the Multiple
lowest cost is found and the corresponding con- Model block diagram function are updated. The
troller is switched in to the closed–loop using the parameters will be stored in to the blocks specified
Multiport Switch. in the Target block. After the parameters have
been stored, the Multiple Model Tuning switch is
V Simulation Example
then switched off, with RLS adaptation switched
The aim of this simulation is to show the difference back on. Then the set–point can be changed again.
of the two different control strategy. The simula- While the simulation is running, all the proper-
tion is separated into two tasks. First, the adaptive ties available in the control Panel can be changed.
control strategy is used. During this period the This is a powerful tool to study adaptive control
model and controller parameters for the Multiple strategy, especially for nonlinear systems.
b) Multiple Model Control [3] K. J. Åström and B. Wittenmark. Self–tuning
controllers based on pole–zero placement. IEE
Once the model parameters for the Multiple Model Proceedings, Part D: Control Theory and Ap-
Control Network have been identified, the cost pa- plications, 127(3):120–130, May 1980.
rameters α, β and λ in the (10) are left to be found.
Parameters α, β and λ are usually chosen empiri- [4] Lennart Ljung. System Identification: Theory
cally, and can be different for each model. Choos- for the User. Prentice Hall PTR, 2nd edition,
ing a different set of parameters for each model Jan. 1999.
can be a difficult task, particularly when the num-
ber of models is large. In this case, parameters α, [5] K. S. Narendra and K. Parthasarathy. Identi-
β were fixed and λ was chosen according to the fication and control of dynamical systems us-
bandwidth of each particular model. As seen from ing neural networks. IEEE Transactions on
Fig. 8 the multiple model approach gives a better Neural Networks, 1(1):4–27, Mar. 1990.
response than self–tuning control. Since an incre- [6] T. Takagi and M. Sugeno. Fuzzy identifica-
mental approach for estimating model parameters tion of systems and its applications to model-
and corresponding controllers was utilized and in- ing and control. IEEE Transactions on Sys-
tegral action was introduced, choosing a different tems, Man and Cybernetics, SMC-15(1):116–
controllers yields a different transient response but 132, Jan.–Feb. 1985.
has no effect on steady–state error.
[7] G. Lightbody M. D. Brown and G. W. Ir-
win. Nonlinear internal model control using
Set−point
0.13
Output local model networks. IEE Proceedings: Con-
0.125
0.12
trol Theory and Applications, 144(6):505–514,
0.115 Nov. 1997.
Ca (mol/l)
0.11
0.095
ple Model Approaches to Modelling and Con-
0.09 trol. Tylor & Francis, London, 1997.
0.085
[1] K. J. Åström and B. Wittenmark. Adaptive [15] Using Simulink. The MathWorks, Inc., Jan.
control. Addison–Wesley Publishing Com- 1999.
pany, Reading, 2nd edition, 1995.