Water Resour Manage

DOI 10.1007/s11269-012-0132-z

Real-Time Operation of Reservoir System

by Genetic Programming
E. Fallah-Mehdipour & O. Bozorg Haddad &
M. A. Mario

Received: 5 March 2012 / Accepted: 27 August 2012

# Springer Science+Business Media B.V. 2012

Abstract Reservoir operation policy depends on specific values of deterministic variables

and predictable actions as well as stochastic variables, in which small differences affect
water release and reservoir operation efficiency. Operational rule curves of reservoir are
policies which relate water release to the deterministic and stochastic variables such as
storage volume and inflow. To operate a reservoir system in real time, a prediction model
may be coupled with rule curves to estimate inflow as a stochastic variable. Inappropriate
selection of this prediction model increases calculations and impacts the reservoir operation
efficiency. Thus, extraction of an operational policy simultaneously with inflow prediction
helps the operator to make an appropriate decision to calculate how much water to release
from the reservoir without employing a prediction model. This paper addresses the use of
genetic programming (GP) to develop a reservoir operation policy simultaneously with
inflow prediction. To determine a water release policy, two operational rule curves are
considered in each period by using (1) inflow and storage volume at the beginning of each
period and (2) inflow of the 1st, 2nd, 12th previous periods and storage volume at the
beginning of each period. The obtained objective functions of those rules have only 4.86
and 0.44 % difference in the training and testing data sets. These results indicate that the
E. Fallah-Mehdipour : O. Bozorg Haddad (*)
Department of Irrigation & Reclamation Engineering, Faculty of Agricultural Engineering & Technology,
College of Agriculture & Natural Resources, University of Tehran, Karaj, Tehran, Iran
e-mail: [email protected]
E. Fallah-Mehdipour
e-mail: [email protected]
M. A. Mario
Department of Land, Air & Water Resources, University of California, 139 Veihmeyer Hall, Davis,
CA 95616-8628, USA
M. A. Mario
e-mail: [email protected]
M. A. Mario
Department of Civil & Environmental Engineering, and Department of Biological & Agricultural
Engineering, University of California, 139 Veihmeyer Hall, Davis, CA 95616-8628, USA

E. Fallah-Mehdipour et al.

proposed rule based on deterministic variables is effective in determining optimal rule curves
simultaneously with inflow prediction for reservoirs.
Keywords Genetic programming . Operational rule curves . Reservoir system . Prediction
model

1 Introduction
In the recent years, limitation of water resources, increasing growth of population, improvement in living standards and increasing growth of industrial regions which causes to increase
consumption, are important reasons for optimal operation of existing resources and structures. A dam stores surface inflow and runoff to avoid impacts of floods and use water to
supply municipal, industrial, and agricultural demands, and hydropower energy generation
in other periods. Operational rule curves are policies that balance reservoir system parameters in each period by considering previous experiences of the system. Those policies
consider physical and hydrological conditions using deterministic and stochastic variables
which are commonly presented in a mathematical equation. Rule curves have been considered by many investigators. Revelle et al. (1969) used a linear decision rule (LDR) in the
design and management of a reservoir. The LDR specified the release during any period of
reservoir operation as the difference between the storage and a decision parameter at the
beginning of the period. Loucks (1970) proposed a LDR including storage and inflow at the
beginning of the period which is regulated by a constant value at each period. Eisel (1972)
developed a chance-constrained reservoir model that employed a more general reservoir
operating policy than earlier chance- constrained models, in which water release was a linear
function of storage, inflow, and satisfied demand at the beginning of the period. Loucks and
Dorfman (1975) considered a model in which water release was related to storage volume
and a weighing factor (a coefficient between zero and one) of inflow in a LDR. Revelle and
Gundelach (1975) proposed an LDR which emphasized the inflow as a stochastic phenomenon. In the proposed rule, released water is a linear function of storage volume at the
beginning of the period and inflows at the beginning and previous periods. Karamouz and
Houck (1982) used dynamic programming and a regression technique as a dynamic
programming regression (DPR) to extract an optimal reservoir operational policy. They
presented a LDR including storage volume and inflow at the beginning of the period which
is regulated with three constant coefficients to achieve an appropriate water release . The
LDRs were used by Moradi-Jalal et al. (2007) and Bozorg Haddad et al. (2009) to extract
optimal reservoir releases as well as cultivated areas in an irrigation system. Bolouri-Yazdeli
(2009) applied first-, second-, and third-order rule curves as linear and nonlinear forms of
rule curves to operate single and multi-reservoir systems. Results showed an improvement in
the value of the objective function with an increase in the rule curve order.
In recent years, most of the investigations addressing reservoir operation rule curves (e.g.,
Oliveira and Loucks 1997; Sharif and Wardlaw 2000; Chang and Chang 2001; Chang et al.
2005a, b; Bozorg Haddad et al. 2008; and Afshar et al. 2011) focused on optimization tools
capable to extract optimal solutions in complex nonlinear problems. Types and capabilities
of optimization tools are important issues to consider in the development of operational rule
curves of reservoir systems. Types of variables and their relations in rule curves are other
important factors in the release of water. As reported in the literature, most rule curves have
linear and nonlinear patterns of inflow and storage volume at the beginning of each period.
The coefficient of this predefined pattern of rule curve has been commonly extracted by

Real-Time Operation of Reservoir System by Genetic Programming

various algorithms. To use these rule curves in real-time reservoir operation, inflow at the
beginning of each period should be calculated as a stochastic variable by a prediction model.
However, water release water can be calculated by using predicted inflow. The latter may be
different from observed inflow and impacts the efficiency of the system. Selection of an
appropriate prediction model is an important issue in the real-time operation of a reservoir
system. Thus, use of an operational rule curve that includes deterministic variables with
inflow prediction at the same time, reduces the amount of calculations in real-time operation
of a reservoir system. This rule curve helps a reservoir operator to calculate water release
without needing a separate inflow prediction model. Thus, an optimization tool with a
considerable capability in optimization and prediction processes is recommended.
GP is a simulation-optimization tool that tries to extract the best simulation (prediction)
relation with a minimum of error between observed and predicted values. Khu et al. (2001),
Rabunal et al. (2007), Sivapragasam et al. (2008), Kisi and Guven (2010), Izadifar and
Elshorbagy (2010), Guven and Kisi (2011), Sreekanth and Datta (2011), Kisi and Shiri
(2011), Azamathulla et al. (2011), and Azamathulla and Ghani (2011) applied GP in realtime runoff forecasting, unit hydrograph determination of a typical urban basin, flood
routing in natural channels, evapotranspiration modeling, hourly actual evapotranspiration
prediction, suspended sediment estimation, coastal aquifer management, precipitation forecasting, development of stage-discharge curve, and prediction of longitudinal dispersion
coefficients in streams, respectively. In all the aforementioned investigations, GP gave
acceptable results compared to other computation tools.
In this paper, a LDR based on the Karamouz and Houck (1982) rule curve is first derived
by genetic algorithm (GA). Two non-predefined pattern rules, including stochastic and
deterministic variables, are then extracted for operation of a reservoir system supplying
downstream demands by GP. Results show the efficiency of the GP rules based on deterministic variables compared to common LDR and GP rules based on stochastic variables.

2 Reservoir System Simulation-Optimization

Reservoir system simulation models calculate reservoir behavior under various conditions
by using a mass balance or continuity equation:
St1 St Qt  Rt  SPt  Losst

in which St and St+1 0reservoir storage volume at the beginning of tth and t+1th periods,
respectively; Qt 0inflow to the reservoir during period t; Rt 0release from the reservoir at
period t; SPt 0volume of spilled water from reservoir at period t; and Lossit 0volume of water
lost from the reservoir at period t.
The simulation models formulation is constrained by the following relations:



Losst F1 Evt ; At

At At At1 =2

At F2 St

E. Fallah-Mehdipour et al.

RMin:  Rt  RMax:

SMin:  St  SMax:

where F1 0function for calculating volume of lost water considering evaporation rate; Evt 0
evaporation depth at period t; At average surface at period t; At 0water surface at the start of
period t; F2 0linear function for transferring storage volume to water surface; RMin., RMax. 0
minimum and maximum allowable capacity for release from reservoir; and SMin., RMax. 0
minimum and maximum storage of reservoir.
In this paper, a dynamic penalty function is used to tackle constraints shown as Eqs. (5)
and (6). The penalty function is added to the minimization objective functions as follows:
PF AfjMin:St  SMin: ; SMax:  St ; 0jg
BfjMin:Rt  RMin: ; RMax:  Rt ; 0jg C

in which PF0penalty function and A, B, and C0positive constants of the linear relation of
the penalty factor.
Although the simulation model shows existing conditions of the system, it is not capable
to determine the best (optimal) condition of the system. To determine an optimal/nearoptimal solution, an optimization model should be coupled with simulation. This model
needs an objective function which tries to achieve its best value. In this paper, the objective
function is considered to be the minimization of the total squared deviation of the released
water for the purpose of meeting downstream demand:

T 
X
Rt  Dt 2
8
Min: Z
Dt
t1
in which: Z0objective function of the supplying downstream demand purpose; T0number
of operational periods; and Dt 0downstream demand of reservoir at period t.

3 Operational Rule Curve Formulation

A rule curve is a mathematical equation which identifies an operational strategy to the
operator by using system experiences. Most of the operational rule curves used as a relation
in simulation-optimization models are linear decision rules (LDR) that include inflow,
storage volume, and released water in each period. Revelle et al. (1969) proposed Eq. (9)
as a rule curve, as follows:
Rt S t  a t

where at 0a constant variable that regulates release water in each period. In contrast,
Loucks (1970) proposed a LDR to operate reservoir system as follows:
Rt St Qt  bt

10

where bt 0a constant variable that regulates storage volume, inflow, and released water in
each period.

Real-Time Operation of Reservoir System by Genetic Programming

Eisel (1972) proposed a LDR which includes demand utility (DUt) as the main factor:
Rt 1  ct St Qt dt  DUt

11

where ct and dt 0constant variables of rule curves.

Loucks and Dorfman (1975) added a weighting factor (t) to the preceding rule:
Rt St lt Qt  et ; 0  lt  1

12

where et 0a constant variable that regulates release water in each period.

Karamouz and Houcks (1982) proposed a LDR to calculate release water in each period:
Rt aSt bQt c

13

in which a, b, and c0constants extracted from the optimization algorithm.

In most of the aforementioned rules, a linear predefined pattern is used in the operation
model with a stochastic variable (Qt) which needs an inflow prediction model so as to be
applicable in real-time reservoir operation.
In this paper, the LDR shown in Eq. (13), which has been widely applied in the literature
(e.g., Mousavi et al. 2007; and Bozorg Haddad et al. 2008), is extracted for the operation of a
reservoir system with the purpose of downstream demand supply. Next, two rule curves
without any predefined pattern which respectively include stochastic and deterministic
variables are derived by GP. In these rule curves, two mathematical functions are calculated:
Rt f1 St ; Qt

14

Rt f2 St ; Qt1 ; Qt2 ; Qt12

15

where f1 and f1 0rule curves that depend on stochastic and deterministic variables, respectively. In Eq. (15), the prediction process is done by using deterministic variables: (1) Qt-1,
(2) Qt-2, and (3) Qt-12 as the inflows of the 1st, 2nd, and 12th previous periods.

4 Evolutionary Computation
Evolutionary computation (EC) is a subfield of artificial intelligence that involves combinatorial optimization. EC uses an iterative process, such as growth or development in a
population to achieve the desired end. Most of this process is often inspired by biological
mechanisms of evolution. EC techniques mostly involve meta-heuristic optimization algorithms such as GA and GP.
4.1 Genetic Algorithm
GA is a search heuristic that simulates the process of natural evolution. This heuristic is
routinely used to generate optimal solutions using techniques inspired by natural evolution,
such as mutation and crossover. In GA, a population of strings which encode candidate
solutions to an optimization problem evolves toward better solutions. The evolution usually

E. Fallah-Mehdipour et al.

starts from a population that is randomly generated in the first generations. In each
generation, the fitness of every solution in the population is evaluated, and the appropriate
solution with the better fitness selected from the current and modified (recombined and
possibly randomly mutated) population to form a new population. Commonly, GA terminates when either a maximum number of generations or iterations is reached, or a satisfactory fitness level is produced for the population. More information about GA is contained in
Goldberg (1989) and Deb (2001).
4.2 Genetic Programming
GP is a type of EC that uses a searching process as in GA. The first statement of GP was
given by Cramer (1985) and later expanded by Koza (1992, 1994). They pioneered the
application of GP in various complex optimization and search problems. GP gives each
solution in a tree structures. Every tree node has an operator function and every terminal
node has an operand, necessitating the evaluation of mathematical and logical expressions.
Figure 1 shows two mathematical samples in the GP. As it is shown, {x,5} and {x,12} are
respectively the terminal sets of yx 15 expx x sinx and yx 2x 12 expressions. The arithmetic operators (, , ) are internal nodes called functions. Also, the
mathematical functions (e.g., sin, cos), Boolean operators (e.g., And, Or), and logical
expression (e.g., If -Then-Else), are identified as the function set. In the GP searching
process, a set of trees is randomly generated and the fitness function of each tree is
calculated. As with most EC, better individuals have a greater chance to survive for the
next generation by using techniques such as roulette wheel, tournament, or ranking method.
The next generation is formulated by using crossover and mutation operators.
In the crossover operator, two parents are selected and sub-tree crossover randomly (and
independently) selects a crossover point (a node) in each parent tree. Then, two children are
produced by replacing the sub-tree rooted at the crossover point in a copy of the first parent
with a copy of the sub-tree rooted at the crossover point in the second parent, as illustrated in
Fig. 2.
In the mutation operator, point mutation is applied on a per node basis. That is, each node
is randomly considered and with a certain probability, it is exchanged by another random
variable, as shown in Fig. 3. The created trees using genetic operators are the input for the
next generation. This process continues up to the termination criterion that may include a
maximum number of generations to be run as well as a problem-specific success predicate,
such as when the error falls below one percent.

Fig. 1 Examples of GP mathematical expressions

Real-Time Operation of Reservoir System by Genetic Programming

Fig. 2 Crossover between parents in GP

5 Application
This paper considers Karaj reservoir as the case study, with an active volume of 176106 m3,
located on the Karaj River with the annual average inflow of 415.23106 m3. This reservoir
has a key role in supplying part of the Karaj and Tehran municipal demands while the other
part is supplied by groundwater. Inflow data are divided into two training (9 years) and
testing (3 years) sets. Figure 4 shows a flowchart of the applied methodology.
At the first step, a LDR which is adapted to Eq. (13) is determined by GA. The best
results (objective) were 35.763 using 50 chromosomes and 100 generations. Table 1 shows
results of five runs of the random-based GA algorithm for LDR and their statistical
measures. Figure 5 shows a decreasing trend of the minimum, average, and maximum
values of the objective function for 5 runs.
At the second step, GP was used for 5 different runs without any predefined pattern
involving 50 trees and 100 generations, the same as the GA function evaluation (50
chromosomes  100 generations). In this rule, storage volume and inflow at the beginning
of the each period were considered in the operational rule curve as in the case of LDR by
GA. Figure 6 shows the decreasing trend of the objective function for GP. Table 1 illustrates
objective function values and their statistical measures for 5 different runs. The best
(minimum) value of the objective function is 12.39 % better than the best value obtained
by GA. Equation (16) is the developed rule with a minimum value of the objective function:
Rt

p
11:351Qt 8:693 sinSt St 6:7042

16

This equation has a nonlinear form, including both arithmetic operators and mathematical
functions.
It should be noted that GA and GP were coded in the software package Matlab7.0 and run
on a PC/WindowsXP/256 MB RAM/2GZ computer with less than one minute execution
time. In addition, four arithmetic operators, including , , and and four mathematical
Fig. 3 Tree structures a before
and b after mutation

(a)

(b)

E. Fallah-Mehdipour et al.
Fig. 4 Flowchart of applied
methodology

Start

Identify initial
population
and algorithm
parameter

Objective function calculation

Selection

Crossover

Mutation

Does rule curve

include stochastic
variables?

Yes

Prediction
model

No
Report realtime operation
results

End

functions involving sin, cos, power (xy), and square root () were considered as the function
set in the applied GP.
Table 1 Results and statistical measures of 5 runs by GA and GP
Method

Optimized objective function values

Statistical measure

Number of run

Minimum

Average

Maximum

Standard
deviation

Coefficient
of variation

GA (LDR)

35.763

35.769

36.110

36.327

36.328

35.763

36.059

36.328

0.282

0.008

GP including
stochastic
variable

31.333

32.234

31.930

31.833

31.948

31.333

31.859

32.234

0.329

0.010

GP including
deterministic
variable

32.863

32.862

33.257

33.895

32.857

32.857

32.147

33.895

0.452

0.014

Real-Time Operation of Reservoir System by Genetic Programming

Fig. 5 Statistical measures
of five runs objective functions
for LDR

90
Minimum
Average

Objective function

80

Maximum
70
60
50
40
30
1

10

19

28

37

46

55

64

73

82

91

100

Generation

This model was solved by language for interactive general optimization (LINGO)
software without considering any operational rule curves and the calculated objective
function value was 27.70, which is 11.59 and 22.55 % less than similar values
obtained by GA and GP, respectively. Both of the derived GA and GP rule curves
include Qt as a stochastic variable which needs a prediction model to apply in the
operation of a real-time system. Wang et al. (2009) used various artificial intelligence
methods to predict inflow with GP, yielding better results than other techniques. In
this paper, GP was employed as the prediction model to be used with rule curves in
the real-time operation of a reservoir. Root mean squared error (RMSE) is assumed as
the objective function and several prediction models based on the different combinations of inflow have been extracted as presented in Table 2. As it is shown, the best
(smallest) values of RMSE were calculated for the training, testing, and total data sets
using Z4. Next, GA- and GP- rules were tested for a 3-year period and the objective
function values of these models were 8.03 and 6.76, respectively. It should be noted
that the objective value of these models without any operational rule curves was 3.57
by the LINGO software. Figure 7 presents release water and storage volume of the
testing data set. As it is shown, more water is released from the reservoir with NLP

Fig. 6 Statistical measures of

five runs objective functions
for GP rule curve including
stochastic variable

70
Minimum
65
Average
Objective function

60

Maximum

55
50
45
40
35
30
1

10

19

28

37

46

55

Generation

64

73

82

91

100

E. Fallah-Mehdipour et al.

Table 2 Comparison of the

inflow prediction model RMSE

Prediction model

Train

Test

Total

Qt Z1 Qt1

18.866

28.597

20.974

Qt Z2 Qt1 ; Qt2
Qt Z3 Qt1 ; Qt2 ; Qt3

14.689
14.905

20.322
19.849

15.863
15.919

Qt Z4 Qt1 ; Qt2 ; Qt12

13.037

18.925

14.289

and GP rule curves. Thus, the storage volume of NLP and GP rule curves is less than
with the GA rule curve in most of the operational periods. In the LDR and GP rules,
released water is a function of the inflow and storage volume. Thus, the variables
changes affect directly released water at each period. In Fig. 7(a), released water by
LDR and GP rules keeps the storage periodic pattern. In contrast, the NLP which
does not follow up any rule pattern and especially a carry-over constraint, and uses
the stored water in the reservoir to supply downstream demands.

(a)
90
LDR

GP

NLP

Release volume (106m3)

80
70
60
50
40
30
20
10
0
1

11

16

21

26

31

36

Period (month)

(b)
250
LDR

GP

NLP

Smin

Smax

200
Storage volume (106m3)

Fig. 7 Volume of a released

water and b storage for the
testing data of the GA- and
GP- rules

150

100

50

0
1

11

16

21

Period (month)

26

31

36

Real-Time Operation of Reservoir System by Genetic Programming

In the next step, a GP rule curve based on deterministic variables that are used in the
prediction model (Z4) and storage volume were extracted as presented in Table 1. The
minimum (best) value of the objective function is 4.86 % more than the best objective value
Fig. 8 Volume of a released
water, b spilled water, and c
storage for the testing data
of GP rule curves

(a)

(b)

(c)

E. Fallah-Mehdipour et al.

by the GP rule base on stochastic variables. Equation (17) is the developed rule with a
minimum value of the objective function:
p
Rt 2:8992Qt1 3Qt2 6:6046Qt12 St 1:7511
17
According to Eq. (16), it is not necessary to employ a prediction model for real-time
reservoir operation and inflow prediction is done with the extracted rule curve.
In the final step, the GP rule curve (Eq. (17)) was tested for a three-year period. The
calculated objective function value was 6.79, showing a 0.44 % difference compared to the
corresponding value using a GP rule based on stochastic variables (6.76). Figure 8 presents
release, spilled water, and storage volume of the testing data set. As it is shown, results of the
two GP rule curves based on stochastic and deterministic variables are similar, with small
differences in some periods.

6 Concluding Remarks
Operational rule curves of a reservoir system are policies to regulate release water for meeting
different operational purposes. These rules include different deterministic and stochastic variables in the form of linear and nonlinear equations. In real-time operation, stochastic variables
should be estimated by using prediction models. The estimated value directly affects the
operational procedures that govern reservoir storage and releases. Inappropriate selection of a
prediction model impacts system efficiency and increases the calculation efforts. Thus, extraction of a rule curve and prediction model at the same time helps to decrease the number of
calculations and probabilistic estimation errors. The aforementioned process needs an appropriate tool with the capability to predict inflow and derive rule curves. In this paper, GP which
presented its capability in inflow prediction, was used as the simulation-optimization tool and
two operational rule curves including water release, storage volume, and inflow were extracted
in the same function evaluation. In the first rule, the stochastic variable (inflow at the beginning
of each period) was estimated by a prediction model for the real-time operation. However, the
inflow at the beginning of each period was predicted using the inflow of the previous period in
the second rule curve. At first, by using GA, the best LDR were calculated and compared with a
GP rule considering stochastic variables. Accordingly, the best (minimum) obtained objective
function was 12.39 % better than the best value obtained by LDR. Then, yields of the GP rule
considering deterministic and stochastic variables were compared in the training and testing
data sets which have 4.86 and 0.44 % differences, respectively. Moreover, there is no considerable difference between storage volume and water release of these rules. Application of GP in
the optimal operation of a reservoir simultaneously with inflow prediction considering other
operation purposes, such as hydropower energy generation, flood control, and even multipurpose operation is recommended for future studies.

References
Afshar A, Shafii M, Bozorg Haddad O (2011) Optimizing multi-reservoir operation rules: an improved
HBMO approach. J Hydroinform 13(1):121139
Azamathulla HM, Ghani AA (2011) Genetic programming for predicting longitudinal dispersion coefficients
in streams. Water Resour Manag 25(6):15371544
Azamathulla HM, Ghani AA, Leow CS, Chang CK, Zakaria NA (2011) Gene-expression programming for
the development of a stage-discharge curve of the Pahang River. Water Resour Manag 25(11):29012916

Real-Time Operation of Reservoir System by Genetic Programming

Bolouri-Yazdeli Y (2009) Effects of extracting multi-reservoir systems operational rule curves for different
reservoirs in system efficiency Master thesis, Department of Irrigation & Reclamation Engineering,
Faculty of Agricultural Engineering & Technology, College of Agriculture & Natural Resources, University of Tehran, Karaj, Tehran, Iran
Bozorg Haddad O, Afshar A, Mario MA (2008) Honey-bee mating optimization (HBMO) algorithm in
deriving optimal operation rules for reservoirs. J Hydroinform 10(3):257264
Bozorg Haddad O, Moradi-Jalal M, Mirmomeni M, Kholghi MKH, Mario MA (2009) Optimal cultivation
rules in multi-crop irrigation areas. Irri and Drain 58(1):3849
Chang L-C, Chang F-J (2001) Intelligent control for modeling of real-time reservoir operation. Hydrol Process
15(9):16211634
Chang F-J, Chen L, Chang L-C (2005a) Optimizing the reservoir operating rule curves by genetic algorithms.
Hydrol Process 19(11):22772289
Chang Y-T, Chang L-C, Chang F-J (2005b) Intelligent control for modeling of real-time reservoir operation,
part II: artificial neural network with operating rule curves. Hydrol Process 19(7):14311444
Cramer NL (1985) A representation for the adaptive generation of simple sequential programs. In: John J (ed)
Proc of an Inter Conf on Genetic Algorithms and the Applications Grefenstette, Carnegie Mellon
University. 2426 July, 18318
Deb K (2001) Multi-objective optimization using evolutionary algorithms. John Wiley & Sons, New York, p 496
Eisel LM (1972) Chance-constrained reservoir model. Water Resour Res 8(2):339347
Goldberg DE (1989) Genetic algorithms in search, optimization, and machine learning Addison Wesley ISBN
0201157675
Guven A, Kisi O (2011) Estimation of suspended sediment yield in natural rivers using machine-coded linear
genetic programming. Water Resour Manag 25(2):691704
Izadifar Z, Elshorbagy A (2010) Prediction of hourly actual evapotranspiration using neural network, genetic
programming, and statistical models. Hydrol Process 24(23):34133425
Karamouz M, Houck MH (1982) Annual and monthly reservoir operating rules. Water Resour Res 18
(5):13371344
Khu ST, Liong S-Y, Babovic V, Madsen H, Muttil N (2001) Genetic programming and its application in realtime runoff forecasting. Journal of Am Water Resour As 37(2):439451
Kisi O, Guven A (2010) Evapotranspiration modeling using linear genetic programming technique. J Irrig
Drain Eng (ASCE) 136(10):715723
Kisi O, Shiri J (2011) Precipitation forecasting using wavelet-genetic programming and wavelet-neuro-fuzzy
conjunction models. Water Resour Manag 25(13):31353315
Koza JR (1992) Genetic programming: on the programming of computers by means of natural selection. MIT
Press, Cambridge MA
Koza JR (1994) Genetic programming II: automatic discovery of reusable programs. MIT Press, Cambridge MA
Loucks DP (1970) Some comments on linear decision rules and chance constraints. Water Resour Res 6
(2):668671
Loucks DP, Dorfman P (1975) An evaluation of some linear decision rules for reservoir planning and
operation. Water Resour Res 11(6):777782
Moradi-Jalal M, Bozorg Haddad O, Karney BW, Mario MA (2007) Reservoir operation in assigning optimal
multi-crop irrigation areas Agri. Water Manag 90(12):149159
Mousavi SJ, Ponnambalam K, Karray F (2007) Inferring operating rules for reservoir operations using fuzzy
regression and ANFIS. Fuzzy Set Syst 158(10):10641082
Oliveira R, Loucks DP (1997) Operating rules for multireservoir systems. Water Resour Res 33(4):839852
Rabunal JR, Puertas J, Suarez J, Rivero D (2007) Determination of the unit hydrograph of a typical urban
basin genetic programming and artificial neural networks. Hydrol Process 21(4):476485
Revelle C, Gundelach J (1975) Linear decision rule in reservoir management and design, 4, A rule that
minimize output variance. Water Resour Res 11(2):197203
Revelle C, Joeres E, Kirby W (1969) The linear decision rule in reservoir management and design, 1,
Development of stochastic model. Water Resour Res 5(4):767777
Sharif M, Wardlaw R (2000) Multi-reservoir system optimization using genetic algorithms: case study. J
Comput Civil Eng (ASCE) 14(4):225263
Sivapragasam C, Maheswaran R, Venkatesh V (2008) Genetic programming approach for flood routing in
natural channels. Hydrol Process 22(5):623628
Sreekanth J, Datta B (2011) Comparative evaluation of genetic programming and neural network as potential
surrogate models for coastal aquifer management. Water Resour Manag 25(13):32013218
Wang W-C, Chau K-W, Cheng C-T, Qiu L (2009) A comparison of performance of several artificial
intelligence methods for forecasting monthly discharge time series. J Hydrol 374(34):294306