19

IWA Publishing 2004

Journal of Hydroinformatics

06.1

2004

Genetic algorithms for real-time operation of multipurpose water resource systems

Alcigeimes B. Celeste, Koichi Suzuki and Akihiro Kadota

ABSTRACT
This paper deals with the application of genetic algorithms to the operation of a water resource system in real time. A genetic algorithm was developed and applied to solve an optimization model for the operation of the system responsible for the water supply of Matsuyama City, in Japan. For comparison purposes, the same model was solved by a technique based on calculus and the Shuffled Complex Evolution Algorithm. The general characteristics of the algorithms and the results from simulations carried out for various conditions are presented. Genetic algorithms appear to be effective tools for real-time reservoir operation.
Key words
Alcigeimes B. Celeste (corresponding author) Koichi Suzuki Akihiro Kadota Department of Civil and Environmental Engineering, Ehime University, 3-Bunkyo-cho, Matsuyama 790-8577, Ehime, Japan Tel./Fax: +81 89 927 9831 E-mails: geimes@yahoo.com ksuzuki@dpc.ehime-u.ac.jp; akado@dpc.ehime-u.ac.jp

| reservoir operation, optimization, genetic algorithms

INTRODUCTION
Ever-increasing water demands, especially in regions where water resources are scarce, require a more efcient management system and a search for ever more elaborate plans for the operation of such systems. There have been a large number of studies concerned with the operation of water resource systems in the literature. They frequently use mathematical models of optimization as tools to identify the best set of policies that provide a fair and economical distribution of the available water. In the water resource eld, linear (LP) and dynamic (DP) programming are the most commonly used methods for the solution of optimization models (Yeh 1985; Mujumdar & Ramesh 1997; Belaineh et al. 1999). Some disadvantages of these two techniques are the need of linear relationships (in the case of LP) or the utilization of various mathematical expressions that have to be simplied in order to reduce their size and the cost of their solution (the so-called curse of dimensionality in DP). A more general formulation and effective handling of the nonlinearities is offered by nonlinear programming (NP). However, the mathematical requirements of NP are more complicated and its standard algorithms are usually based on calculus and are not always efcient for complex
Figure 1

problems containing non-convex, discontinuous and non-differentiable functions. Genetic algorithms (GAs) are search techniques that have been shown to overcome the limitations of the conventional calculus-based algorithms for nding the global (or near-global) solutions of nonlinear optimization problems in which the surfaces

Location and layout of the study system.

20

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

usually nd only local solutions. Since GAs do not require a continuous surface for computing gradients, discontinuous objective functions do not create problems. Although there have been many applications of genetic algorithms in water resources, its use for optimal operation is not frequent (Wardlaw & Sharif 1999). This study aims to investigate the applicability of genetic algorithms to solve an optimization model for the operation of a water resource system in real time. The well known Shufed Complex Evolution Algorithm (SCE-UA) (Duan 1991) and a Sequential Quadratic Programming (SQP) method were also utilized to solve the model and comparisons were carried out. The study system is responsible for the water supply of Matsuyama City, the capital of Ehime Prefecture, which lies on the northern part of the island of Shikoku, the smallest of the four major islands of Japan. Matsuyama has a mild climate with an average temperature of 15.6C and an average annual precipitation of 1333 mm. There is much rain in June but only a little in January and the city receives little snow and few typhoons. The system consists of a multipurpose reservoir called the Ishite River Dam and a set of 26 wells located around the Shigenobu River (Figure 1). The reservoir controls the overow of the Ishite River and provides half of the water supply of Matsuyama while the underground water of Shigenobu River provides the other half. The dam is also used for irrigation of the northern area of Ishite River. The main objective is to determine the allocation of water from the reservoir and the set of wells that best solve the demands for water supply and irrigation. Another aim is to maintain the reservoir storage as close as possible to a given target storage in order to not let it decrease too much. Besides, the distribution of water should not compromise the operation of the system, violating its constraints and leading it to collapse.

Figure 2

Outline of the tank model used and its parameters.

METHODS
The optimization model To formulate a mathematical optimization model we need an objective function and usually, but not necessarily, a set

of the objective functions are noisy and rough (Goldberg 1989; Michalewicz 1996; Michalewicz & Fogel 2000). For such kinds of surfaces, gradient-based methods can

21

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

120

Efficiency = 97% Correlation = 98%

100

100

200 80

Runoff (mm)

300

60

400

500 40 600 20 700

0 1 31 61 91 121 151 181 211 241 271 301 331 361

800

Days
Rainfall

Obs. Runoff

Cal. Runoff

Figure 3
48 47 46 45

Results from calibration of the tank model.

0 100 200 300 400

44 43 500 42 41 40 39 38
May-95 Mar-95 Mar-96 Aug-95 Aug-96 Jun-96 Oct-96 Oct-95 Apr-96 Nov-96 Dec-96 Nov-95 Dec-95 Jan-95 Jun-95 Feb-96 Feb-95 Jul-96 Jul-95

600 700 800 900

Efficiency = 97% Correlation = 99%

Date
Rainfall
l. Elevation
Obs. Elevation

Figure 4

Results from calibration of the multiple linear regression model.

Table 1

Target values of demands and reservoir storage utilized JulySeptember: 160,000 m3/d 1

of constraints. As shown in Figure 1, the three main decision variables of the system are: Qrelallocation of water from the reservoir for city supply; Qwellallocation of water from the wells for city supply and Qirramount of water allocated from the reservoir for irrigation. Based on the objectives, and considering an operating horizon of N days, the mathematical model for the operation may be written as shown by the expressions (1)(7). The objective function is formed by the sum of deviations
Target reservoir storage Demands for city supply

Rainfall (mm)

Elevation (m)

Rainfall (mm)

OctoberJune: 140,000 m3/d 1

MaySeptember: 19,000 m3/d 1 Demands for irrigation OctoberApril: 7,000 m3/d 1

7,500,000 m3

22

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 5

(a) and (b).

23

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 5

Optimal results from the SQP approach. (a) Reservoir storage and optimal releases, (b) city supply, (c) irrigation.

of the desired releases and storage volumes from their targets. Constraints are given by the continuity equation, limitations of the components of the system, etc. The problem thus lies in nding the values of the decision variables that minimize the losses resulting from deviations from their target values: minimise
N

(3) Water provided by the wells depends on the precipitation (P) and cannot be greater than the capacity of the underground water treatment plants (Qmax): well Qt #f(Pt); t = 1, . . ., N well Qt #Qmax; t = 1, . . ., N well well (4) (5)

t l

HS

t t Qrel Qwell{Tt dem

Tt dem

DS DS
2 t t Qirr{Tirr t Tirr 2

t Vstor{Tstor

Tstor

DJ
2

(4) Water released from the reservoir cannot be greater (1) than the capacity of the surface water treatment plants (Qmax): rel Qt #Qmax; t = 1, . . ., N rel rel (5) Nonnegativity constraints: Qt $0; Qt $0; Qt $0; t = 1, . . ., N rel well irr (7) (6)

subject to (1) Mass balance/continuity equation for the reservoir: Vt + 1 = Vt + Vt Vt stor stor inf evap Qt Qt Vt ; t = 1, . . ., N (2) rel irr spill (2) Storage volume (Vstor) can vary only between the maximum (V reservoir: V
dead stor max stor dead stor

) and the dead storage (V

) of the

where t = day, Tt dem is the desired water amount for city supply (target demand) on day t, Tt is the desired water irr amount for irrigation on day t, Tstor is the desired storage at the reservoir, Vt is the expected inow to the reservoir inf

#V

t stor

#V

max stor

; t = 1, . . ., N

(3)

24

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 6

(a) and (b).

25

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 6

Optimal results from the SCE-UA approach for 1996. (a) Reservoir storage and optimal releases, (b) city supply, (c) irrigation.

on day t, Vt evap is the water evaporated from the reservoir on day t and Vt spill is the water spilled from the reservoir on day t.

1. 2.

Set T = 1 (T = 1, 2, . . ., Ttot) and specify the initial reservoir storage (e.g. target storage). Forecast the rainfall for the next N days. Use the forecasted rainfall to estimate the inows to the reservoir and the amount of water available in the wells for the next N days.

Real-time operational procedure Since the operation of the system is to happen in real time, the allocation of water must be decided without perfect knowledge of future events. Consequently, the operational procedure is formulated to accept forecasted rainfall data over a period of N days. The model (1)(7) is thus solved to nd the optimal water distribution for an N-day operating horizon. However, only optimal releases for the rst day of the operating horizon are implemented. The releases determined for the other days will be adjusted during the next model run as updated forecast information is obtained. Considering T, the index of the current operating day, and Ttot, the total number of days of operation, the steps in the simulation process are as follows: 4. 5. 6. 3.

Solve problem (1)(7) and obtain the optimal reservoir releases and amount of water to be drawn from the wells during the next N days. Implement the rst days releases/withdrawals. Update the reservoir storage using the actual reservoir inow and the selected releases. Set T = T + 1. If T > Ttot stop; otherwise go to step (2).

The estimation of the inows to the reservoir and the amount of water that would be available in the wells for a given operating horizon (step 2) were obtained using two approaches: (1) use of a rainfallrunoff model to evaluate the inows and (2) use of a multiple linear regression

26

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 7

(a) and (b).

27

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 7

Optimal results from the SCE-UA approach for 1996 under 80% of actual rainfall. (a) Reservoir storage and optimal releases, (b) city supply, (c) irrigation.

model to correlate precipitation with average water level in the wells and subsequently determine the amount of water. The following subsections give a brief explanation of these methodologies.

Estimation of the water available in the wells To predict the amount of water that would be available in the wells over a given operating horizon, the following multiple linear regression model was tted to the historical data of precipitation and average water level of the wells:

Estimation of the inflows to the reservoir In order to estimate the daily inows to Ishite River Dam from the precipitation, the rainfallrunoff Tank Model (Sugawara 1995) was calibrated for its watershed by using the developed genetic algorithm. A four-tank model with fourteen parameters was used (Figure 2). The calibration was based on the minimization of the sum of squares of the differences between the observed and simulated daily discharges. Figure 3 displays the comparison of observed and simulated runoff for the calibration period. A detailed description of this calibration can be found in Celeste (2002).

EL(i) = fEL(i 1) + lP(i 1) + gP(i) + O

(8)

in which EL is the average daily water elevation in the wells, P is the average daily precipitation, f, l, O and g are the model parameters and i is the index of the day. Therefore, the volume of water in the wells for the days in the operating horizon could be calculated by using the average water level estimated from the forecast rainfall (Celeste 2002). The results from the calibration of model (8) are shown in Figure 4.

28

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 8

(a) and (b).

29

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 8

Optimal results from the SCE-UA approach for 1995. (a) Reservoir storage and optimal releases, (b) city supply, (c) irrigation.

Overview of the constructed genetic algorithm In order to solve problem (1)(7), a genetic algorithm was implemented in MATLAB. GAs apply heuristic rules in search of the best solution for a given optimization problem. They mimic Darwins evolution process by implementing a survival of the ttest strategy. In principle, the search starts with a set of initial random solutions called population. Individuals, called chromosomes, compose the population. Each chromosome represents a potential solution for the problem and is described by a string of symbols (usually, but not necessarily, a binary bit string). Every solution is evaluated to provide some measure of tness. A new population is then formed by selecting the more-t individuals. Some members of this new population are submitted to transformations by means of genetic operators (crossover and mutation) to form new solutions, called offspring. After successive iterations, or generations, the algorithm converges to the best chromo-

some, which hopefully represents the optimum or suboptimal solution to the problem. In general, the more-t individuals tend to reproduce and thus improve the successive generations. However, inferior individuals may happen to survive and also reproduce (Michalewicz 1996; Gen & Cheng 1997; Houck et al. 1995). It has been shown that a more natural representation of the chromosomes is more efcient and yields better solutions than the classical binary implementation. Michalewicz (1996) did extensive experiments and showed that oating point coding is faster and provides higher precision than binary representation. The genetic algorithm developed in this work uses oating point representation for the chromosomes and incorporates stochastic universal sampling as the selection procedure. Stochastic Universal Sampling (SUS), proposed by Baker (1987), is a selection method where the population is represented by a pie chart, in which each zone is associated with a chromosome of the population.

30

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 9

(a) and (b).

31

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 9

Optimal results from fGENEC for 1996 (a) Reservoir storage and optimal releases, (b) city supply, (c) irrigation.

The areas of the zones are proportional to the tness of their representative chromosomes. A number of equally spaced pointers equal to the population size is put around the pie and after a single wheel spin the chromosome pointed to by each marker is selected for recombination (crossover and mutation). This approach is better than roulette wheel, apparently the most popular selection procedure, because it keeps the diversity and prevents superchromosomes from dominating the population. Six genetic operators, as used in Michalewiczs GENOCOP system (1996), were also employed:

Non-uniform mutation: randomly selects one variable and sets it equal to a non-uniform random number; Arithmetic crossover: produces two complementary linear combinations of the parents (chromosomes selected for recombination); Simple crossover: The head (rst sequence of values up to a certain position randomly chosen) of one individual is connected to the tail (remaining sequence after the crossover position) of the other, and vice versa, to produce two new offspring;

Uniform mutation: randomly selects one variable k and sets it equal to a uniform random number from the range (Lk,Uk), the lower and upper bound, respectively, for variable k;

Heuristic crossover: produces a linear extrapolation of two parents utilizing tness information.

One characteristic of model (1)(7) is that, although its objective function is nonlinear, all its constraints are either linear (a set of restrictions dened by expression (2)) or domain constraints (inequalities (3)(7)).

Boundary mutation: a variation of uniform mutation that selects a random variable and sets it equal to either Lk or Uk;

32

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 10

(a) and (b).

33

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 10

Optimal results from fGENEC for 1996 under 80% of actual rainfall. (a) Reservoir storage and optimal releases, (b) city supply, (c) irrigation.

When the set of constraints is linear, the offspring generated by the above operators are always feasible, considering (Lk,Uk) to be the feasible range of a variable k where other variables remain xed (Michalewicz 1996). This feature is then sufcient to handle the constraints of problems (1)(7). The evaluation function used to determine the tness of each chromosome in the population is, in principle, the objective function of the optimization problem. However, after the function is evaluated for all the chromosomes, the population is submitted to what is called tness ranking, a process where individuals are sorted in order of raw tness (actual values of the objective function), and then reproductive tness values are assigned according to linear ranking. In the case of minimization, the smaller the objective function value the bigger the rank and, consequently, the tness. Because of this, the effect of one or two extreme individuals will be negligible in the selection process, independent of how much greater or less their

tness is than the rest of the population (Beasley et al. 1993a, 1993b). Another feature of the algorithm is the use of a procedure called elitism, according to which the best and worst points of the new population are detected and the worst one is replaced with the best chromosome from the old population. Elitism avoids the best chromosome being lost from one generation to another by crossover or mutation.

Other techniques used Apart from the above-described GA (which will be from now on referred to as fGENEC, where f and C designate oating point representations of the chromosomes and constrained optimization, respectively), the optimization problem (1)(7) was also solved by two other techniques. The rst one was a calculus-based procedure, which belongs to the MATLAB Optimization Toolbox and uses a Sequential Quadratic Programming (SQP) method. In this

34

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 11

(a) and (b).

35

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Figure 11

Optimal results from fGENEC for 1995. (a) Reservoir storage and optimal releases, (b) city supply, (c) irrigation.

method, a Quadratic Programming (QP) subproblem is solved at each major iteration. A full description of this procedure is found in the toolbox manual (Optimization Toolbox Users Guide 1996). The second one was the well known Shufed Complex Evolution Algorithm (SCE-UA), originally developed by Duan as part of his doctoral dissertation work at the University of Arizona (Duan 1991). In the SCE-UA method, the search starts from a population of points sampled randomly from the feasible space. The population is partitioned into one or more communities (complexes), each containing a xed number of points. Each complex evolves, based on a statistical reproduction process that uses the simplex geometric shape (Nelder & Mead 1965) to lead the search in an improved direction. Periodically, the entire population is shufed and points are reassigned to complexes to ensure information sharing. As the search progresses, the entire population tends to converge towards the neighborhood of the global opti-

mum, provided the initial population size is sufciently large.

RESULTS
A detail that must be stated is that the real-time use of the model requires inputs that are actual forecasts of rainfall. However, such data are not available for past years. Thus, all rainfall forecasts were assumed to be perfect and obtained by using actual measured rainfall data. All the runs of the model were performed for one year of operation and an operating horizon of ve days. The target values of demands and reservoir storage were chosen as in Table 1. The capacity of the dam was assumed to be up to the ood control level, which means a volume of 8,500,000 m3. The dead storage of the reservoir is 2,200,000 m3.

36

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

(a)

9,000,000 8,000,000 7,000,000

operation with the same characteristics as the one used for testing the SQP procedure. In Figure 7, an operation for 1996 using only 80% of the actual rainfall and an initial reservoir storage equal to that observed was performed. Figure 8 presents the results for one operation in 1995 with the initial reservoir storage equal to the observed volume.

Reservoir Storage (m3)

6,000,000 5,000,000 4,000,000 3,000,000 2,000,000 1,000,000 0 1 31 61 91 121 151 181 211 241 271 301 331 361

Results from fGENEC The same set of simulations done for the SCE-UA was performed by fGENEC. Figure 9 shows the results from an operation with the same characteristics as the one used for testing the SQP procedure. In Figure 10, an operation for 1996 using only 80% of the actual rainfall and an initial reservoir storage equal to that observed was performed. Figure 11 presents the results for one operation in 1995 with the initial reservoir storage equal to the observed volume.

Days
fGENEC

SCE-UA

(b)

9,000,000 8,000,000 7,000,000

Reservoir Storage (m3)

6,000,000 5,000,000 4,000,000 3,000,000 2,000,000 1,000,000 0 1 31 61 91 121 151 181 211 241 271 301 331 361

Days
fGENEC
SCE-UA

Figure 12

Reservoir storage over the year of operation by the GA-based approaches for (a) 1996 and (b) 1995.

DISCUSSION
We can observe from the results regarding the same scenario (Figures 5, 6 and 9) that the SQP approach has

Results from the SQP approach Figure 5 shows the results from an operation using the calculus-based SQP technique. This operation uses the observed data of the year 1996 as perfectly forecast rainfall. In this run, the initial reservoir storage was assumed to be equal to the maximum reservoir storage considered (8,500,000 m3). Figure 5(a) displays the behavior of the reservoir storage and the optimal releases over the year. Figure 5(b) considers only the results for city supply and shows how much the total allocation from the reservoir and wells achieves the target demands. Figure 5(c) regards the results for irrigation and also shows how the releases attain (or do not attain) the demands.

decided not to release enough water to meet the demands during certain days (e.g. days 5878 for the city supply) while the SCE-UA and fGENEC found that the demands could still be satised on those days without compromising the system. This fact suggests that the GAs have found lower values of objective function than the SQP for such periods. Since the objective function is quadratic we could have expected the results using SQP to be better compared to the GAs. This would be undoubtedly ensured if we were dealing with an unconstrained problem. Nevertheless, our problem contains a large set of constraints which might have caused difculties to the SQP technique in nding a global solution. In fact, the MATLAB Optimization Toolbox warns us that the procedure to solve quadratic programming problems may nd only local minima, depending on the constraints. Another point to be mentioned is that the use of the real-time operational procedure may also have inuenced

Results from the SCE-UA For the SCE-UA as well as fGENEC, several simulations were carried out. Figure 6 shows the results from an

37

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

the performance of the algorithms. Each gure presents the results for a whole year of operation but the solution for each day is obtained by solving the model only for that day plus four days forward. Thus, the operational year is not considered completely in a single run, which in fact cannot be done because of the use of forecast information. The operational process in real time thus has the features of a greedy algorithm that executes a sub-algorithm (in this case SQP, fGENEC or SCE-UA) which performs the optimization for ve days and uses only the rst one. Hence, the solution is not optimal (which could be obtained only by taking the whole year into consideration) but a sub-optimal one. The GA-based approaches then seem to achieve better solutions than the SQP when performing as the sub-algorithm of the greedy-like operational procedure. For the operation in 1996 using both the SCE-UA and fGENEC (Figures 5, 6 and 9) the demands for city supply were not attended to between days 155 and 168 mainly because the water available in the wells during those days was very low. Even though the reservoir would have had enough water to supply the demands, this water could not be released since the amount needed would be greater than the capacities of the surface water treatment plants. For the operation in 1995, when the initial reservoir storage was assumed to be that observed (Figures 8 and 11), the demands for city supply could not be reached during few days in April since the reservoir volume would decrease signicantly and thus the releases from the dam should not be very high. Between the two approaches based on GAs, fGENEC has provided slightly better results than the SCE-UA because it was able to keep more water in the reservoir, whereas the demands were met in nearly the same way as in the operations by the SCE-UA. This is seen in Figure 12, for example. It has to be mentioned here that the SCE-UA was performed using its default parameters and improved results could have been achieved by selecting other combinations of parameters. Several other tests under different scenarios were investigated and the results obtained from them were similar to the above solutions (Celeste 2002).

CONCLUSIONS
Because of the limitations of the conventional calculusbased techniques for global optimization of discontinuous and multimodal functions there has been an increase in popularity of random search algorithms based on natural laws, which do not require the functions to be continuous, convex and differentiable. The genetic algorithm is perhaps the most widely known example of such methods. There have been many applications of genetic algorithms in the water resource eld but only a few of them are related to the optimal operation of reservoirs. For this task, linear and dynamic programming are most prevalent. This study aimed to examine the application of genetic algorithms to solve an optimization model for the operation of a multipurpose system in real time. A GA based on oating point representation of the chromosomes and composed of various operators was developed and implemented for the resolution of the problem. Its performance was compared with the SCE-UA method and a technique based on Sequential Quadratic Programming. The simulations carried out illustrated that evolutionary approaches are also efcient and robust for applications concerned with the operation of water resource systems.

ACKNOWLEDGEMENTS
The eld data were provided by the Matsuyama Construction Works Ofce. The SCE-UA algorithm was developed by Dr Qingyun Duan at the Department of Hydrology and Water Resources, University of Arizona.

REFERENCES
Baker, J. E. 1987 Reducing bias and inefciency in the selection algorithm. In: Proceedings of the Second International Conference on Genetic Algorithms, pp. 1421. Lawrence Erlbaum.

38

Alcigeimes B. Celeste et al.

Genetic algorithms for real-time reservoir operation

Journal of Hydroinformatics

06.1

2004

Beasley, D., Bull, D. R. & Martin, R. R. 1993a An overview of genetic algorithms: Part 1, fundamentals. Univ. Comput. 15(2), 5869. Beasley, D., Bull, D. R. & Martin, R. R. 1993b An overview of genetic algorithms: Part 2, research topics. Univ. Comput. 15(4), 170181. Belaineh, G., Peralta, R. C. & Hughes T. C. 1999 Simulation/ optimization modeling for water resources management. J. Wat. Res. Plann. Mngmt. 125(3), 154161. Celeste, A. B. 2002 Optimal real-time operation of a multipurpose water resource system using genetic algorithms. MSc thesis Department of Civil and Environmental Engineering, Ehime University, Matsuyama, Ehime, Japan. Duan, Q. 1991 A global optimization strategy for efcient and effective calibration of hydrologic models. PhD dissertation Department of Hydrology and Water Resources, University of Arizona, Tucson, AZ. Gen, M. & Cheng, R. 1997 Genetic Algorithms & Engineering Design. Wiley-Interscience, New York. Goldberg, D. E. 1989 Genetic Algorithms in Search, Optimization and Machine Learning. Addison Wesley, Reading, MA.

Houck, C., Joines, J. & Kay, M. 1995 A genetic algorithm for function optimization: a Matlab implementation. NCSU-IE Technical Report 95-09, North Carolina State University. Michalewicz, Z. 1996 Genetic Algorithms + Data Structures = Evolution Programs, 3rd edn. Springer-Verlag, New York. Michalewicz, Z. & Fogel, D. B. 2000 How To Solve It: Modern Heuristics. Springer-Verlag, New York. Mujumdar, P. P. & Ramesh, T. S. V. 1997 Real-time reservoir operation for irrigation. Wat. Res. Res. 33(5), 11571164. Nelder, J. A. & Mead, R. 1965 A simplex method for function minimization. Comput. J. 7, 308313. Optimization Toolbox Users Guide Version 5 1996 The MathWorks Inc., Natick, MA. Sugawara, M. 1995 The development of a hydrological model-tank. In: Time And The River (ed. G. W. Kite), pp. 201257. Water Resources Publications. Wardlaw, R. & Sharif, M. 1999 Evaluation of genetic algorithms for optimal reservoir system operation. J. Wat. Res. Plann. Mngmnt. 125(1), 2533. Yeh, W. W.-G. 1985 Reservoir management and operations models: a state-of-the-art review. Wat Res. Res. 21(12), 17971818.