Multireservoir System Optimization

Using Multi-objective Genetic Algorithms

Taesoon Kim1 and Jun-Haeng Heo2
1

PhD Candidate, School of Civil and Environmental Engineering, Yonsei University,

Seoul, 120-749, South Korea; email: chaucer@yonsei.ac.kr
2
Professor, School of Civil and Environmental Engineering, Yonsei University,
Seoul, 120-749, South Korea; email: jhheo@yonsei.ac.kr

Introduction
Multireservoir system consists of several reservoirs which are connected serially
or parallel in the same basin. To optimize such a complex multireservoir system, the
dynamic programming (DP), linear programming (LP) and non-linear programming
(NLP) have been widely applied to real problems. However, when DP is applied to
multireservoir system it has a major problem, so called the curse of dimensionality
and LP and NLP have essential approximation problems dealing with discontinuous,
nondifferentiable, non-convex, or multi-modal objective functions. Recently, there
has been an increasing interest in a biologically motivated adaptive system for
solving optimization problems. The genetic algorithms (GAs) are one of the most
promising techniques in natural adaptive system field and receiving many attentions
because of their flexibility and effectiveness for optimizing complex systems.
Optimization of multireservoir system is to solve multi-dimensional and multiobjective problems and GAs are appropriate optimization methods to multireservoir
system. GAs are not restricted by a number of dimensions because computer memory
increases by dimensions linearly, not exponentially. Thus, there is no curse of
dimensionality. Especially classical optimization methods such as DP, LP, and NLP
are not proper to multi-objective optimization because these methods use a point-bypoint approach, in which the outcome of classical optimization methods is a single
optimal solution. However, GAs use a population of solutions in each iteration
instead of a single solution, so they are called as population-based approaches. This
is one of the most striking differences between classical optimization methods and
GAs.
In this paper, multi-objective GAs are applied to optimize multireservoir system

of the Han River basin in South Korea. Multi-objective GAs, which have many
attractive features, have had only limited applications to the multireservoir system
optimization. The present work focuses on the application of multi-objective GAs to
the multireservoir system optimization. The solutions of multi-objective GAs yield a
trade-off curve or surface, identifying a population of points that define optimal
solutions to the problem. Non-dominated sorting approach is used to get the nondominated fronts and maintaining a diverse set of solutions in the non-dominated
fronts is achieved by sharing. Crossover and mutation operators are used and
tournament selection is applied. Chromosomes are coded by real values.

Multi-Objective Genetic Algorithms

Multi-Objective Genetic Algorithms (MOGAs) are a part of GAs. It has
chromosome, population and basic three operators such as reproduction, crossover
and mutation. The different thing between GAs and MOGAs is a ranking. MOGAs
have two objectives at least and these multiple objectives should be conflicting
objectives like cost vs. effect. The ranking is which chromosome dominates the
others in terms of its performance on conflicting objectives and non-dominated
solutions with same rank make a trade-off curve.
The ranking approach used in this paper is a fast-non-dominated-sort in NSGA-II
(Deb et al., 2002). In the first iteration of ranking, chromosomes are given as a rank
of 1 if there are no chromosomes which perform better and said to be non-dominated
solutions in the first iteration. After assigning ranks to all members of chromosome
in population, the top ranking chromosomes (in this case rank of 1) are removed. In
the second iteration, all the remaining chromosomes are assigned to their own ranks
and the non-dominated solutions have rank of 2. The non-dominated solutions from
each iteration make fronts of non-dominated solutions and represent trade-off curves.
A number of researchers have developed different implementation of MOGAs.
Among these, Multiobjective Genetic Algorithm (Fonseca and Fleming, 1998a;
Fonseca and Fleming, 1998b), Nondominated Sorting Genetic Algorithm (Srinivas
and Deb, 1994) and Strength Pareto Evolutionary Algorithm (Zitzler and Thiele,
1999) are most imitated MOGAs (Van Veldhuizen and Lamont, 2000). Fonseca and
Fleming (1998a, 1998b) and Van Veldhuizen (2000) examine major MOGAs
techniques. Recently, MOGAs are applied to different real-world problems (Cheng
and Li, 1997; Cieniawski et al., 1995; Vink and Schot, 2002; Yeh and Labadie, 1997).

Model Formulation
Initial Population
Generally, initial population of GAs is randomly calculated between upper and
lower bounds of each variable. However, in reservoir operation optimizations,
randomly calculated initial population must have a large amount of infeasible
solutions. Because releases are serially and highly connected to the front and back,
although the release of preceding month is feasible, it does not guarantee that all the
remaining next releases are feasible.
To overcome this problem, initial population is calculated by considering current
storage and inflow of each reservoir. If the current storage and inflow are sufficient,
the probability that release has large value is increasing. In contrast, if the current
storage and inflow are insufficient, release should be calculated to small value. Using
this method, more feasible solutions can be calculated.

Crossover
Simulated Binary Crossover (SBX) is applied to crossover. The procedure of
calculating the children solutions ( c1 , c2 ) from the parent solutions ( p1 , p2 ) is as

follows (Deb and Goyal, 1996).

A uniform random number ( u ) between 0 and 1 is generated and using Eq. (1)
is computed.
the spread factor
1

(2u ) n+1
=

1
2(1 u )

if u

0.5
(1)

1
n +1

otherwise

in which n is a distribution index of SBX and any nonnegative real number.

Then, the children solutions are calculated as follows.
c1 = 0.5 [(1 +

) p1 + (1

) p2 ]

c 2 = 0.5 [(1

) p1 + (1 + ) p 2 ]

(2)

These two children solutions are symmetric about the parent solutions. A larger
value of the distribution index ( n ) allows that the children solutions are closer to the
parent solutions. A smaller value of n makes a more uniform distribution in the

range 0

1 and if n equals to 0, then it makes a uniform distribution in the

same range.

Fast-Non-Dominated-Sort and Sharing

Fast-Non-Dominated-Sort (FNDS) approach is used to assign a rank to
chromosome (Deb et al., 2002). FNDS has a computational complexity of
O( MN 3 ) (in which M is the number of objectives and N is the population size)
and a sharing parameter to preserve diversity of solutions is calculated by crowding
distance of solutions. The best solution of population is passed to next generation by
elitism.

Objective Functions
In this study, two objective functions are considered, because three or more
objective functions are used, most non-dominated solutions have rank of 1 so fast
and then there is little improvement of objective function values. The first objective
is the sum of storage and the second one is the difference between release and water
supply. Eqs. (3) and (4) are the objective functions.
1st Objective Function : Min i

2nd Objective Function : Min -

Current Storage(i, t)
Maximum Storage(i, t)

(3)

W [Release(i, t ), Supply(i, t )]

(4)

Where i is reservoir location ( i = 1, 4 ), t is period ( t = 1, 12 ), W is a weighting

function. Eq. (3) is the proportion of current storage and maximum storage and Eq.
(4) has a larger value (maximum value is 1) as release is getting closer to water
supply.

Application
Han River Basin
The Han River consists of the North and South Han Rivers and has three large
reservoirs such as Hwacheon, Soyanggang, and Chungju. The total storage of
Hwacheon is 1,022MCM (million cubic meters), Soyanggang is 2,900MCM and
Chungju is 2,750MCM. These three reservoirs are located at upstream in parallel and
Paldang reservoir, located at the confluence of the North and South Han Rivers, is
used as a control point (Fig. 1). Hwacheon, Soyanggang, and Chungju reservoirs

play an important role on water supply and flood control to downstream, especially
Seoul Metropolitan area.

GAs Parameters
Chromosome consists of 36 real values which are the monthly releases of three
reservoirs. The probability of crossover is 0.9 and that of mutation is 1/36. Generally,
the performance of MOGAs is evaluated by the performance of crossover, which is
the SBX in this study. Thus, the probability of crossover is higher than usual case.
The number of chromosomes in a population is 500 and iteration number is 500. It
takes about 5~10 minutes with Pentium 4 3.2Ghz computer to finish simulation once.

Hwacheon

Chuncheon

Soyanggang

Uiam

Chungpyong

Chungju

Storage Reservoir

Paldang

Flow-Through Reservoir

Fig. 1. Schematic location of reservoirs

In the Han River Basin

Distribution Index
SBX uses the distribution index ( n ) to control how much closer the children
solutions are to the parent solutions. A large value of n makes the children solutions
very close to the parent and a small value of n allow distant points to be selected as
children solutions. Deb and Kumar (1995) showed that a large distribution index
( n =150) is appropriate to multimodal problems and they used the value of n =30 in
multi-objective problems. First of all, we tried to use the value of n =30 in our

multi-objective optimization problems. But, the releases are changed abruptly as

some releases are much small and the others are much large. In real-world, decision
maker prevent releases from changing rapidly, because the abrupt change of releases
is easy to cause flooding or water shortage.
In order to find a proper value of n, the distribution indexes from 2.0 to 5.0
increased by 0.1 are tested, because the single-point crossover is successful to find
good building blocks if it is used with proper coding schemes and the performance of
SBX with real-coded GAs is similar or better than that of the single-point crossover
in the range of 2~5 (Deb and Agrawal, 1995). As a result, SBX with distribution
index 3.0 and 3.8 show the best results. The results used other distribution index
values show a large amount of water shortage, especially at Choongju reservoir in the
last period.

Analysis of Pareto-optimal solutions

Pareto-optimal solutions are a group of non-dominated solutions with rank of 1.
If multiple objectives used in simulation are conflicting objectives with each other
and selected well, Pareto-optimal solutions form trade-off curve. But the objectives
are selected poorly, Pareto-optimal solutions are gathered into one point or small
region. In this case, the objective functions should be selected again.
In this paper, we used two objective functions. The first objective function is the
total sum of proportional storage of each reservoir and the second one is the sum of
weighted difference between release and water supply and Pareto-optimal solutions
make a trade-off curve well (Figs. 2 and 3). The upper-left corner solutions (A) have
larger storage and smaller water supply. On the other hand, the lower-right corner
solutions (C) have larger water supply and smaller storage. The solutions near B are
neutral solutions between the first and the second objectives.
In order to examine the decision space, a point is selected from near A, B, and C,
respectively in each distribution index ( n =3.0 and 3.8). The water shortages and the
coefficient of variation of these selected points are presented in Table 1. The negative
water shortages means that releases are not enough to water supply. Coefficient of
variation (VAR) is calculated to discriminate which month has the power to improve
objective function values. At Hwacheon reservoir, only the last month has high VAR
about 81.82 (n=3.0) and 84.48 (n=3.8) and the other months have low values. A
decision maker, who knows these results, tries to find good reservoir operating
policies at Hwacheon with changing releases of last month, September. VARs of

Soyanggang reservoir are sensitive to the distribution index. In n = 3.0, the VAR has
high values in four months (December, May, June, and September) but in n = 3.8, it
has only one high value in September. Choongju reservoir has three or four high
VARs over 10.0 in both distribution indexes.
-22

2nd Objective Function

-24

-26

-28

-30

-32
-35.52 -35.41 -35.29 -35.18 -35.06 -34.95 -34.83 -34.72 -34.60 -34.49 -34.37

1st Objective Function

Fig. 2. Pareto-optimal solutions (n = 3.0)

-22

2nd Objective Function

-24

-26

-28

-30

-32
-35.39 -35.27 -35.14 -35.02 -34.89 -34.76 -34.64 -34.51 -34.39 -34.26 -34.14

1st Objective Function

Fig. 3. Pareto-optimal solutions (n = 3.8)

Table 1. Water shortages of Pareto-optimal solutions (million cubic meters)

Water shortage (n = 3.0)

Dam /
Month

VAR

Water shortage (n = 3.8)

A

VAR

HC

10
11
12
1
2
3
4
5
6
7
8
9

-4.6
-5.9
-3.7
-3.9
0.5
-3.9
-5.5
288.0
360.2
150.7
-12.9
-110.0

-2.3
-4.8
-1.2
-4.1
0.8
-4.0
-2.7
264.0
366.0
161.3
-4.9
-4.1

-4.4
-4.3
-1.3
-3.5
0.5
-2.9
-2.0
292.9
366.6
137.9
-1.7
-0.8

2.94
1.30
2.82
2.05
0.73
0.38
4.94
4.48
0.68
3.21
1.80
81.82

-6.1
4.1
-4.8
-2.0
1.6
-2.9
2.2
348.4
254.6
176.8
-0.3
-118

-2.9
-3.4
-0.5
-1.0
-1.3
1.3
-0.4
315.3
317.7
147.0
-28.8
-8.7

-3.0
-2.3
0.6
-1.8
-0.9
-0.1
-0.3
310.2
321.4
161.1
0.9
2.4

2.43
7.28
4.22
0.89
3.35
5.69
2.52
4.38
8.15
4.23
8.45
84.48

SY

10
11
12
1
2
3
4
5
6
7
8
9

-10.5
-4.6
-27.7
-8.1
-5.1
-3.2
-9.9
239.9
7.9
414.7
-50.0
-119.5

-6.3
-5.9
-3.0
-6.3
-2.8
-2.6
0.8
-0.8
258.0
324.7
-2.5
3.1

-1.1
-2.6
-0.7
-8.0
-1.7
-2.7
-1.8
-0.7
265.0
328.4
-2.0
3.3

1.81
0.56
12.01
0.71
1.09
0.15
4.27
55.34
49.86
9.83
9.10
38.96

-7.8
-2.8
-14.1
-10.9
-2.7
-8.3
-7.1
-0.4
334.8
278.3
-17.6
-117.5

-6.1
-6.5
-8.4
-4.7
-1.3
-10.5
3.2
-2.0
375.7
215.1
-3.1
-15.4

1.0
-1.4
-5.6
-4.5
-0.4
-4.3
1.4
-0.3
372.4
276.1
1.8
-2.0

2.07
1.12
3.22
2.63
0.91
1.17
3.79
0.86
4.31
6.84
5.48
34.80

CJ

10
11
12
1
2
3
4
5
6
7
8
9

-15.4
-21.6
-41.9
-17.8
-184.3
-13.5
-8.3
-1.7
178.8
1209.1
387.7
-279.1

-4.6
-4.8
-4.7
-12.7
-175.8
-5.8
-5.7
1.2
-4.2
1353.5
337.6
0.4

-0.3
-3.0
-2.2
-8.9
0.4
-0.8
-3.7
-0.6
-5.1
1207.7
340.4
0.6

1.40
3.56
7.94
1.07
82.84
1.59
0.58
0.47
24.07
4.70
3.60
113.67

4.1
-8.3
-29
-6.9
-221.6
-56.8
-12.4
-2.0
119.7
1292.5
393.7
-208.1

-7.3
-10.2
-10.4
-11.0
-153.9
-5.7
-10.1
1.5
-1.5
1288.2
393.0
-278.2

-1.9
-3.8
-1.8
-7.3
0.5
-5.1
-7.8
1.4
-0.7
1281.4
281.8
1.3

1.99
0.88
5.04
1.78
98.51
10.21
0.59
0.44
16.09
1.02
8.35
110.59

Note: (1) HC (Hwacheon reservoir), SY (Soyanggang reservoir), CJ (Choongju reservoir); (2) VAR
means coefficient of variation of releases from total optimal solutions.

Summary and Conclusions

The research reported in this paper focused on the application of multi-objective
genetic algorithms (MOGAs) to reservoir operation optimization. The developed
MOGAs are applied to multi-reservoir system in Han River basin in South Korea.
The objective functions consist of two objectives which are the sum of storage and
the difference between release and water supply. Crossover rate is 0.9 and mutation
rate is 1/36. The number of chromosome and the iteration number are 500 and 36
real-coded variables are used to make chromosome. The ranking approach is fastnon-dominated-sort of NSGA-II and simulated binary crossover (SBX) is used. The
distribution indexes ( n ) of SBX are set to 3.0 and 3.8.
MOGAs have been demonstrated to be an effective solution technique for solving
multireservoir system optimization. The approach can efficiently identify Paretooptimal solutions (a trade-off curve) that exist for a multi-objective optimization
problem. The trade-off curve can be used by a decision maker to obtain an
appropriate solution considering the conflicting objectives which are to maximize the
storages and minimize the water shortages.
Some improvements over the algorithm used in this study are possible. Firstly,
the releases are correlated to the next stage serially, thus if crossover and mutation
operator with correlation are developed, the performance of MOGAs could be
improved. Another future work could explore the application of MOGAs involving
more than two objectives. To manage more than two objectives, a sub-ranking
mechanism, which can sort chromosomes within same rank and prevent
chromosomes from premature, is necessary.

References
Cheng, F. Y., and Li, D. (1997). "Multiobjective optimization design with pareto
genetic algorithm." Journal of Structural Engineering, 123(9), 1252-1261.
Cieniawski, S. E., Eheart, J. W., and Ranjithan, S. (1995). "Using genetic algorithms
to solve a multiobjective groundwater monitoring problem." Water Resources
Research, 31(2), 399-409.
Deb, K., and Agrawal, R. B. (1995). "Simulated binary crossover for continuous
search space." Complex Systems, 9, 115-148.
Deb, K., and Goyal, M. (1996). "A combined genetic adaptive search (GeneAS) for

engineering design." Computer Science and Informatics, 26(4), 30-45.

Deb, K., and Kumar, A. (1995). "Real-coded genetic algorithms with simulated
binary crossover: Studies on multimodal and multiobjective problems." Complex
Systems, 9, 431-454.
Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. (2002). "A fast and elitist
multiobjective genetic algorithm: NSGA-II." IEEE Transactions on Evolutionary
Computation, 6(2), 182-197.
Fonseca, C. M., and Fleming, P. J. (1998a). "Multiobjective optimization and
multiple constraint handling with evolutionary algorithms - Part I: A unified
formulation." IEEE Transactions on Systems Man and Cybernetics Part aSystems and Humans, 28(1), 26-37.
Fonseca, C. M., and Fleming, P. J. (1998b). "Multiobjective optimization and
multiple constraint handling with evolutionary algorithms - Part II: Application
example." IEEE Transactions on Systems Man and Cybernetics Part a-Systems
and Humans, 28(1), 38-47.
Srinivas, N., and Deb, K. (1994). "Multiobjective optimization using nondominated
sorting in genetic algorithms." Evolutionary computation, 2(3), 221-248.
Van Veldhuizen, D. A., and Lamont, G. B. (2000). "Multiobjective evolutionary
algorithms: Analyzing the state-of-the-art." Evolutionary Computation, 8(2), 125147.
Vink, K., and Schot, P. (2002). "Multiple-objecttive optimization of drinking water
production strategies using a genetic algorithm." Water Resources Research,
38(9), 20-1-20-15.
Yeh, C.-H., and Labadie, J. W. (1997). "Multiobjective watershed-level planning of
storm water detention systems." Journal of Water Resources Planning and
Management, 123(6), 336-343.
Zitzler, E., and Thiele, L. (1999). "Multiobjective evolutionary algorithms: A
comparative case study and the strength Pareto approach." IEEE Transactions on
Evolutionary Computation, 3(4), 257-271.

10