Beruflich Dokumente
Kultur Dokumente
Pbrez-Correa*,
Francisco
Several exact linearization methods were applied to a simplified non-linear model for the concentration of
dissolved oxygen in a waste water treatment plant, and digital control algorithms were derived based on
these linearizations. A non-linear adaptive control algorithm is proposed and compared with a well tuned
PID, a linear adaptive controller, and a non-adaptive non-linear controller. The proposed algorithm shows
a better performance under a variety of perturbations. However, users must be careful in the choice of the
appropriate model parameters to be estimated.
(Keywords: models; adaptive control; non-linear algorithms; dissolved oxygen; water treatment)
There are many non-linear systems in the process industry with a well known qualitative behaviour. These
systems can be modelled by non-linear dynamic equations obtained from first principles. However, modelling
includes several empirical parameters (kinetic constants,
transport coefficients, etc.) that are difficult to estimate
and vary unpredictably. The usual approach to deal with
this type of system in process control is to use local
linearization and then apply linear control theory to
design the controller. Processes with weak non-linearities
can be properly controlled in this way, if the control
design is robust enough. However, in highly non-linear
processes (pH neutralization, non-isothermal reactors),
control design based on linear theory usually gives an
undesirable performance.
A different approach, which has been shown to be very
successful, consists of incorporating
the non-linear
model of the process directly into the control calculations. This approach is often called model based control
in the process literature. A closely related technique uses
an exact linearization of the non-linear model of the
process to design a control algorithm in the linear transformed space. Several methods have been proposed to
find appropriate transformations;l-9 among these, the
geometric methods are most widely used. Unfortunately,
most of these papers assume perfect model knowledge
and present only simulation results. For an overview of
geometric approaches applied to process control, the
reader is referred to the works of KantorO, Henson and
Seborg, McLellan et al.*, Isidorii3 and Kravaris and
Kantori4.
152
0959-1524~91~030152-09
0 1991 Butterworth-Heinemann
Ltd
$2 [G(t)
KdCs
- y(t)] + K,u(t)[Cs(t)
K2UNN
- ~1
The controllability index (r) can be defined as the smallest integer that satisfies:
(dh, a4-(g))
where (dx,y)
# 0
(3)
adz0t_d = y
(44
4 Cv)= kl
WI
adzZ01) = [z,[z,~ll
The Lie bracket is defined by:
[Z,Yl
ay
&Z
a2
ZY
(2) is 1,
= Kl[Cs - y] # 0
Pd.vldt+ POY= v
(7)
Control algorithms
dY) =
Plvw
(8)
&)lu
Toh
Yk[PI
POT01
(9)
- y(t)] (1)
vk
!L!yx+ +[Ci
-yk] - K&b,)
~=fi)+&)u
(24
hb9 =
(2b)
(10)
with
+ Kl[Cs - yk]uk
[PI
- yJ/Ti
(11)
153
~Y(z) -
(21)
Y*
p.9 -
T&c]z-2
[Kc/ ToZ]z-
+ [KCT, + (Kc/Ti) -
CS,cV)(dgldY) = a
(22a)
Cf,ot,u)(dgldY)
(22b)
Kc = [0.9 - P,P,]/T,
(13)
(14)
(1%
Yk)]
= bv
(23)
(24)
K(Cs -
are used:
1.9]z- + 1 (12)
Yk
(25)
&Vk
Variable transformation4
(16)
Gfd_d + WX_YJ~)
(17)
(184
GfiW = a
Cf2~,u)
dgldt = a + bv
(19)
154
uk
[ is-_;;
]uk _ , +
W,(Cs-
yk)]
(26)
This is the same result given by the GLC method (Equation (15)), where:
Kc = [0.1/T,, + Kc]
(27a)
Ti = KcT,z/[1.9
(273
(28)
where fi,u,d,t)
is a model of the process. If the nonlinear model given by Equation (1) is used, the following
algorithm is obtained:
~[C@)
(29)
-KIUk(CS.-Yk)-~CCYk-Yk+I)+~2[LOlk)-~Cyk+,ll=O
7.6~
Rearranging the above equation results in the same controller obtained by the GLC method (Equation (15)).
The same control algorithm can be derived using any
of the three basic methods described above. The GMC is
simpler to use; however, as noted by Henson and
SeborgO, it is not clear which conditions ensure that a
solution to the GMC problem exists.
Estimation
The above control algorithms require the estimation of
the model parameters. This can be done by discretizing
the process model, and then expressing the difference
equation in a form suitable for recursive parameter estimation. Depending on the data used in the estimation, a
positional or an incremental algorithm can be derived.
Positional estimation
The model Equation (1) can be discretized to give:
~k+I=Yk+TO[~k(Cik-Yk)/~+~I~k(C~-Yk)--K21
(31)
and rearranging:
or in an equivalent way:
=
yk+l
Yk
tTo&J'kl/V
&To
@kTo)/u
(33)
+'kek
where:
+'k
[UkTo(cs
wk
[K,,
yk),
ci,
Yk-I
[T,Qk-,Yk-d/V
&To
+'k-lek-1
(34)
AYk -
&t&Y/c
Qk-IYk-d/V
@'k
ok
(35)
where:
wk
[To(csAuk
Ah_Yk)),
To@,
Qk-d/q
A recursive estimation algorithm can be used with Equation (35) in the same way as with Equation (33).
Results
Several simulation runs were carried out. The perfor-
4.44
0
20
Time
110 0
80
60
40
(min)
control
(0)
vs non-
155
4.0
0
-L--l-II
8
Time
1
12
16
1
20
l
24
6.95
1
0
12
Time
(min)
Figure 3
Measured
Table 2
load changes.
16
20
24
(min)
PID (0),
adaptive
GLC ( x )
changes
Time (min)
Ci (mg 1-l)
3.75
9.25
15.25
22.96
5.0
0.5
1.5
3.0
0
0
12
b
Figure 2
b, input
Table 1
Time
PID (0),
16
20
24
[min)
adaptive
GLC ( x ). a, Output,
Q (1 min-I)
3.75
9.25
15.25
22.96
0.070
0.020
0.060
0.040
156
I
4
I
8
Unmeasured
12
Time
Figure 4
1.51
disturbances.
I
16
I
20
I
24
(min)
PID (0),
adaptive
GLC ( x )
:
i
0.4-i
0
II
4
11
8
11
12
11
16
1
20
I
24
6.8
12
Time (min)
Figure 5
Table 3
16
Time
Figure 6
Variations
-0.2
III
20
24
(min)
3.15
9.25
15.25
22.96
3.3
1.4
2.1
1.9
e
0
III
4
12
Time
II
16
I
20
1)
I
24
(min)
The PID controller presented a stable, but slow, response (not shown). On the other hand, the self-tuner was
faster (Figure 8a), but the control action was more sensitive (Figure 86) and the controller gain suffered some
variations (Figure 9). A good estimation of the mass
transfer coefficient is very important for the GLC algor-
Table 4
K, (1-I)
3.75
9.25
15.25
22.96
1.30
1.80
0.92
0.60
157
71
0
I
16
12
Time
I
20
I
24
6.73
0
12
(mini
Time
Figure 10 Variations
in K,. Adaptive
estimated. Output ( x )
Figure 8
a, Output,
Variations
b input
kl
l2 (min) l6
Time
(0),
20
16
24
[min)
GLC,
only
K, parameter
is
1
P
IL
2o
adaptive
24
GLC ( x ).
12
Time
Figure 11
1 .l
Real parameter
(0),
16
20
24
(mini
estimated
parameter
(x)
I
I
l.Ok
II
-0.1
I
4
12
Time
Figure 9
gain (Oh
Normalized estimated
K, (x ), Ci (A)
16
20
24
(min)
parameters.
Linear self-tuner
control
15%
6.701
12
Time
Figure 12
tion ( x )
Variations
20
16
24
(mini
in K,. Adaptive
estima-
2.0
1.5-
.a,
g
u9
l.O-
Acknowledgement
VI
2
5
0.5-
:
2
0
III
4
Time
Figure 13
Estimated
parameter
12
(-).
16
II
20
I
24
References
(mini
Real parameter
(x)
Conclusions
A non-linear adaptive algorithm was developed to
control the dissolved oxygen concentration in a waste
water treatment plant. A simplified non-linear model was
employed to derive the controller and to test it through
simulations. It was shown that different linearizing techniques can give the same control algorithm; however, the
GMC approach is simpler to use. An adaptive version of
the more general GLC controller was compared with
different standard controllers under a variety of disturbances. In this particular example, the PID gave a reasonable response to all disturbances, but was comparatively slow and oscillatory. A minimum time linear selftuner always showed better behaviour than the PID.
However, model parameter changes caused undesirable
variations in the estimated controller gain, which sometimes took on the opposite sign. The non-adaptive GLC
with nominal parameters can cope efficiently with a wide
range of .disturbances; however, it is very sensitive to
variations in the mass transfer rate. In this situation,
adaptation is called for.
There is no unique way to design an adaptive GLC
controller. It is necessary to specify which model para-
159
Nomenclature
;
d
C,, C*
Ci
CS
160
arbitrary constant
arbitrary constant
disturbance
constants
inlet oxygen concentration (mg 1-I)
saturation concentration (mg 1-l)
%!u,d,r)
f%?
t??
h(z)
k
K,
K2
Kc
L(x)
p,, p*
Q
r
TO
Ti
V
Y
Y
Y
z
2
w. PI
!J
CD