Documentos de Académico
Documentos de Profesional
Documentos de Cultura
12 International Conference on
Environmental Science and Technology
CEST2011 Rhodes, Greece
8-10 September 2011
EXTENDED ABSTRACT
Normally, Reverse Osmosis desalination plants (RO plants) include one or two PID
controllers. A particular control problem with such plants consists in the fact that plant
parameters change fast because of fouling and therefore, membrane cleaning has to
carried out often (e.g. once a week). Thus, process parameters obtained after cleaning
are very different from the parameters obtained one week later before a new cleaning
task. Hence, the control performance deteriorates very quickly in the course of the week,
when the controller was optimal adjusted by using one particular model. Moreover, the power
supply based on renewable energies is affected by fluctuations due to weather conditions.
These fluctuations are also reflected in operating point changes and consequently yield
parameter uncertainties.
One way of dealing with parameter uncertainties is to apply a robust control approach.
However, plant manufacturers do not like to use other control strategy than PID (Proportional-Integral-Derivative). In the current work, the PID controllers for permeate flow rate
of a RO plant powered by renewable energies are designed by means of a multi-objective
parametric optimization approach so that the control loops are less sensitive to parameter
changes of the plant.
The multi-loop control system is designed by using dynamic input-output models whose
parameters are obtained before and after cleaning the membranes. As performance indices,
ISE (Integral Square Error) for each model and controller are used. All performance
indices are optimized simultaneously with the aid of NSGA-II (Non-dominated Sorting
Genetic Algorithm, version two). The final outcome is obtained from the Pareto front
according to the Nash solution of a bargaining game.
Simulation results show that the proposed method has a better performance in a wide range
of operation conditions than controllers designed for one particular model. Next research
steps are the design of an adaptive control system in order to compare its performance
with the robust control approach and the real-time implementation, which will allow testing
the control system design in the real laboratory plant.
Keywords: reverse osmosis, robust control, multi-objective optimization
1.
INTRODUCTION
Reverse osmosis desalination plants use sensible components, which are also prone to
parameter changes because membranes are sensitive to temperature of feed water,
fouling, scaling, pressure variations and changes in the power supply because of
variations in the weather conditions. RO plants are normally controlled by using PID
control laws, which are tuned but not optimized.
A particular control problem with RO small plants consists in the fact that plant parameters
change very fast because of fouling and membrane cleaning has to carried out often (e.g.
once a week). Thus, process parameters obtained after cleaning are very different from
the parameters obtained one week later before a new cleaning. Therefore, the control performance deteriorates very fast in the course of the week, when the controller was optimal
adjusted by using one of these models. Hence, a robust control approach should be used but
plant manufacturers do not like to use other control strategy than PID control.
In the present work, an optimal PID control system for a reverse osmosis desalination laboratory plant is presented. The methodology consists in obtaining a unique set of controller
parameters for a family of models of the plant by means of multiobjective optimization
(MOO). In Section 2, the RO plant and its dynamic characteristic are described. Section 3
is devoted to present the methodology for the controller design. In Section 4, the MOO
method and the decision making procedure are described, so that simulation results are
shown and analyzed is Section 5. Finally, conclusions are drawn in Section 6.
2.
C F T
T P
Filter
Permeate
HP
Pump
P
F
Brine
Control
signal 1
Control
signal 2
Valve 1
Valve 2
Permeate
flow rate
Brine
flow rate
Bypass
flow rate
RO
Plant
Permeate
conductivity
=
B # ( s ) [-139.26, -1.77] s + [-56.51,-2.3875]
(2)
and
A# ( s ) =
s 3 + [0.71, 0.96]s 2 + [0.112, 0.635]s + [0.003, 0.071] .
(3)
The nominal model is randomly obtained and presented for information but later it is not used
for the controller design. This nominal model is
B( s) =
29.791 s 39.172]
(4)
A( s ) =
s 3 + 0.776s 2 + 0.287s + 0.0492 .
(5)
and
Dynamic properties obtained from the models are shown in Figure 3.
200
Bode Diagram
Amplitude
20
Imaginary Axis
Root Locus
100
0
10
-100
-200
Phase (deg)
Magnitude (dB)
180
-10
0
-180
1000
-1
-2
10
10
Imaginary Axis
10
Frequency (rad/sec)
10
10
Nyquist Diagram
-20
-200
-100
0
Real Axis
100
200
Step Response
160
140
500
120
100
80
60
40
-500
20
-1000
-500
Real Axis
Lower model
500
Upper model
50
100
Nominal model
150
200
250
Time (sec)
Figure 3: Process characteristics for the minimum, maximum and nominal models
3.
sTD
1
( s ) K b r ( s ) y ( s ) +
[r ( s) y ( s )]
u=
y(s) .
1 + sTD / N
sTI
(6)
where normally 0 b 1 and 8 N 20. Eq. (6) leads to a polynomial description given
by
P=
( s )u ( s )
T ( s)r ( s) Q( s) y ( s) ,
(7)
where
P ( s ) = p0 s 2 + p1 s + p2 ,
(8)
T ( s ) = t0 s + t1 s + t2 and
(9)
Q ( s ) = q0 s 2 + q1 s + q2 .
(10)
=
p0 1,=
p1 N / Td ,=
p2 0 ,
(11)
t0 ==
b K , t1 K / Ti + b K N / Td , t2 =
K N / (Ti Td ) ,
(12)
(13)
Finally, parameters of the PID controller can be calculated back by using the following
formulas
(q1 p1 q2 ) / p12 , b =
K=
t0 / K , Ti =
K p1 / q2
(14)
N =q0 / K 1 and Td =N / p1
(15)
A( s ) y ( s ) = B ( s ) u ( s )
(16)
and the control error is defined as e(s) = r(s) y(s), then the closed loop transfer function
from r(s) to e(s) and to u(s) are
e( s ) [ A( s ) P( s ) + [Q( s ) T ( s )]B( s ) ]
=
r (s)
A( s ) P( s ) + B( s )Q( s )
(17)
u (s)
A( s )T ( s )
,
=
r ( s ) A( s ) P ( s ) + B ( s )Q( s )
(18)
and
respectively.
3.3. Controller design
The controller design follows the steps proposed in Gambier (2009) for the Integral of the
Square Error (ISE) as performance index, which is defined by
J= J e + J u ,
where
(19)
2 j
Ju =
e (t ) dt
=
=
Je
u 2 (t ) dt =
e( s )e( s ) ds ,
(20)
s 2u ( s )u ( s ) ds
(21)
2 j
and is a free weighting parameter. The model of the process be now defined as interval
polynomials denoted by
A# ( s ) y ( s ) = B # ( s ) u ( s )
(22)
(23)
(24)
and
with m < n, then the PID controller is obtained by using the lower and upper polynomials
defined by
(25)
(26)
J in J e ,in + J u ,in .
=
(27)
=
J e,in
2 j j
ein ( s )ein ( s ) ds
(28)
with
ein ( s ) =
uin ( s ) =
and
Ain ( s )T ( s )
r ( s) .
Ain ( s ) P( s ) + Bin ( s )Q( s )
(29)
(30)
Evolutionary algorithms can be applied to solve SOO problems as well as MOO problems.
The MOO case was first studied in Goldberg (1989). Today, there are many MOEAs
distinguished mainly by the algorithms for the population ranking in the fitness
assignment. The most important are: MOGA (Multiple Objective Genetic Algorithm,
Fonseca, 1995), NSGA-II (Non-dominated Sorting Genetic Algorithm, Deb et al., 2000),
SPEA2 (Strength Pareto Evolutionary Algorithm, Zitzler et al).
A very important advantage of the MOEAs is that they do not need information about the
objective-functions derivatives and they can solve non-convex problems. Moreover, the
algorithms are relatively robust and they do not require solving a sequence of singleobjective problems and it is a parallel search technique. However, Pareto optimality is not
a concept embedded in the fundaments of evolutionary algorithms. Consequently, it could be
that a Pareto optimal solution born and then, by chance, also it dies out. An additional
drawback is the intensive computational burden required until the solution is obtained. NSGAII is a very efficient evolutionary algorithm that can be used for solving off-line problems land
therefore it is chosen for the present study.
The decision making can be done by using bargaining games (Thomson, 1994).
For example (see Thomson, 1994), the Nash solution of the game (NS) corresponds
to a point of the Pareto set which yields the largest rectangle (c, B, NS, A), (Figure
4), the Kalai-Smorodinsky solution (KS) is situated at the intersection of the Pareto
front and the straight line, which connects the threat point and the utopia point,
and the egalitarian solution (ES) yields the point given by the intersection of the
Pareto front and a 45-line through the threat point. The controllers parameters
obtained by means of NSGA II and NS selector are summarized in Table 1.
Ju
Threat
0.53
0.529
0.528
0.527
0.526
CS
NS KS
ES
0.525
0.524
Utopia
0.646
0.647
0.648
0.649
0.65
0.651
0.652
0.653
Jl
t0
t1
p1
q0
q1
q2
CS
6e-3
-0.4
6.64
-0.13
-0.22
-4e-3
NSGA-II
KS
NS
6e-3
6e-3
-0.4
-0.4
6.71
6.66
-0.13
0.13
-0.22
-0.22
-4e-3
-4e-3
ES
6e-3
-0.4
6.66
-0.13
-0.22
-4e-3
5.
SIMULATION RESULTS
The approach described above has been studied in a simulation environment. The
operating point was set to 250 l/h for a valve opening of 60%. At this point, the flow rate
and the valve opening are assumed to be zero. Hence the reference is changed to 385
l/h (i.e. 35 l/h over the set point) and after 20 second the set point is changed again to the
value of 270 l/h (20 l/h regarding the operating point). This experiment is carried out first
considering first the optimal controller obtained by the NBI algorithm and the lower and upper
models. Moreover, an additional model is obtained randomly from the parameter intervals.
Simulation results are shown in Figure 5. As it is possible to observe the controller performs
satisfactory for all three models.
40
30
20
10
0
-10
0.4
0.2
0
-0.2
-0.4
-0.6
0
5 10 15
u1 valve opening
20
25
20
25
Lower model
10
15
Upper model
30
35 t0 [s]
30 35 t0 [s]
Random model
Figure 5: Permeate flow rate and control signals for the optimal PID
controller and several models of the plant
The second study consists in analysing the behaviour of the controllers obtained for the
absolute minimums, i.e. Jlo,min and Jup,min. In order to simplify Figure 6, the model obtained
in the random way is not considered in this case.
40
30
20
10
0
-10
0.4
0.2
0
-0.2
-0.4
-0.6
0
5
10 15
u1 valve opening
20
25
30
35 t0 [s]
0
5
10 15 20 25 30 35 t0 [s]
PID for Jup,min with lower model
PID for Jlo,min with upper model
PID for Jup,min with upper model
PID for Jlo,min with lower model
MOO-PID
Figure 6. Example of mismatch between controller and model used for the design
Figure 6 shows that controllers, which are optimized for a particular model, do not perform
correctly in the case of model mismatch. Thus, RO plants present some control
difficulties, which can be solved by using multi-objective optimization.
6. CONCLUSIONS
In this work, the control problem of a reverse osmosis desalination plant, which parameters
are uncertain is solved by using the multi-objective NSGA II algorithm, when the plant can
be modelled by means of interval polynomials.
Due to the fact that plant manufacturer wish to use PID controllers, this algorithm was
used for the design. The controller requires a MOO in order to obtain a unique set of
coefficient for the polynomials Q, T and P.
Simulation results shown the good performance of the controller for several models,
which are randomly chosen in the region defined by the intervals polynomials.
REFERENCES
1.
2.
3.
4.
5.
6.
7.
8.
9.
strm, K. J., and T. Hagglund, Advanced PID Control. ISA - The Instrumentation, Systems, and
Automation Society, 2005.
Gambier, A., Optimal PID controller design using multiobjective Normal Boundary Intersection
th
technique. Proc. of the 7 Asian Control Conference 2009, 1369-1374, Hong Kong, August 27-29,
2009.
Deb, K., S. Agrawal, A. Pratab and T. Meyarivan. A fast elitist non- dominated sorting genetic
algorithm for multi-objective optimization: NSGA-II. In: Lecture Notes in Computer Science,
(M. Schoenauer, K. Deb, G. Rudolph, X. Yao, E. Lutton, J. J. Merelo, and H-P. Schwefel,
Eds.), Springer, No. 1917, 849-858, 2000.
Fonseca, C. M., Multiobjective genetic algorithms with application to control engineering problems. Ph.D.
Thesis, University of Sheffield, Sheffield, U.K., 1995.
Goldberg, D. E. Genetic algorithms in search, optimization, and machine learning. AddisonWesley, Reading, Massachusetts, 1989.
Kunkle, D., A summary and comparison of MOEA algorithms. Internal Report, College of Computer
and Information Science Northeastern University, 2005.
Thomson, W., Cooperative models of bargaining. Handbook of Game Theory with Economic App.,
in: R.J. Aumann & S. Hart (ed.), 2, 1237-1284, Elsevier.
Zitzler E., M. Laumanns and L. Thiele. SPEA2: Improving the Strength Pareto evolutionary algorithm.
Tech. Report 103, CH-892 Zurich, Switzerland, 2001.
Zitzler, E., K. Deb and L.Thiele, Comparison of multiobjective evolutionary algorithms: Empirical results.
Evolutionary Comp. Journal, 8, 125148, 2000.