Professional Documents
Culture Documents
Application of Genetic Algorithm For The Groundwater Management of A Coastal Aquifer
Application of Genetic Algorithm For The Groundwater Management of A Coastal Aquifer
Application of Genetic Algorithm For The Groundwater Management of A Coastal Aquifer
To cite this article: S. K. Pramada, S. Mohan & P. K. Sreejith (2017): Application of genetic
algorithm for the groundwater management of a coastal aquifer, ISH Journal of Hydraulic
Engineering, DOI: 10.1080/09715010.2017.1378597
Download by: [La Trobe University] Date: 30 September 2017, At: 02:19
ISH Journal of Hydraulic Engineering, 2017
https://doi.org/10.1080/09715010.2017.1378597
pumped for domestic consumption. The calibrated groundwater model was used as the simulation
component in the linked simulation–optimization model. The combined model was then used to
identify the optimum pumping strategy.
1. Introduction into the computer simulation model process can provide sig-
nificant performance improvement over the results obtained
In nature, physical processes are often represented by large sets of
by traditional stand-alone simulation or optimization meth-
nonlinear, partial differential equations. Unfortunately, this fact
odology. For applying the simulation–optimization method-
limits the ability of many traditional optimization codes to solve
ology, the first step is the development of a simulation model
water resources management problems and leads to necessary
to compute the head values at different locations for a given
simplification of the actual problem. An alternative approach for
pumping strategy. In second step, the simulation model needs to
solving complex problems is to employ a combined simulation– be externally linked with the optimization model. Whenever the
optimization approach. In combined simulation–optimization optimization procedure requires the objective function and or
approach, the output from the simulation model forms the input constraint evaluation, it calls the simulator and passes the modi-
to the optimization model. The significant advantage of inter- fied input parameters to the simulator and the simulation model
facing a simulation model with an optimization algorithm is executes and does the required evaluation. Genetic algorithm
that the solution requires no additional simplifying assump- module acts as a driver model where in it calls the simulation
tions about the physics of the real system. The simulation model model by passing the management decision variables and gets
can be combined with the optimization model either using the back the corresponding objective function value. The genetic
governing equations as binding constraints in the optimiza- algorithm then adjusts the management decision variables to get
tion model (embedding technique) or using a response matrix a new objective function value and continue the process until
method or by response surface method (Bhattacharjya and there is no further improvement in the objective function by
Datta 2005, Dhar and Datta 2009, Esfahani et al. 2016, Khosrow altering the decision variables. The genetic algorithm toolbox in
and Hasan 1996, Pramada et al. 2014). Matlab 7 (The MathWorks 2003) was used as the optimization
SEAWAT (Langevin et al. 2003), one of the widely used coastal routine, in which, it calls the calibrated simulation model during
aquifer simulation models (Holzbecher 1998, Mahesha and each iteration. A subroutine has been developed to interface
Lakshmikant 2014, Motz and Sedighi 2013) could be combined the genetic algorithm tool box with the calibrated model (to
with heuristic optimization techniques such as genetic algo- communicate between both modules) using Matlab.
rithm to solve optimization-based groundwater management Usually interfacing a simulation and optimization model
problems. Owing to the complexity of the data files required by means making a communication possible with each module.
the SEAWAT establishing an interface between the simulation That is whenever an optimization models requires a function
model and optimization models is not an easy procedure. evaluation or constraint evaluation, it calls simulation model.
The batch run utility in the simulation model was used to cou-
ple the optimization model with the simulation model. Figure 1
2. Simulation–optimization
shows the program execution sequence. The program of genetic
A combined simulation–optimization approach coupling algorithm starts from an initial guess solution given by the
groundwater simulation model with optimization routine based user. Pumping rates were generated randomly by the MATLAB
on GA is proposed. This integration of optimization technique random number generator and GA performs generations of
Numerical model
Run
Downloaded by [La Trobe University] at 02:19 30 September 2017
Run
Model.BAT
No Stop? Yes
Print RESULTS
possible solutions. For every set of random pumping rates, many input files, but during pumping rate optimization, only
SEAWAT is executed once to update the hydraulic head dis- the pumping rate of individual well changes. All other param-
tribution in response to those pumping rates. The output of eters, for example hydraulic conductivity, specific storage, etc,
the simulation consists of the values of the hydraulic heads do not change, hence only the ‘well’ file is modified (contains
at all model nodes/cells. The values of hydraulic heads for the details pertaining to pumping) using the interfacing program
selected observation points are forwarded to the optimization developed using Matlab during each iteration. The idea here is
module. The optimization algorithm performs the evaluation to introduce new pumping rate through the ‘well’ file and then
of the constraints and checks for termination, and if the con- run the batch file. Check the file containing the head obser-
straints are not satisfied, then it computes new pumping rates vations and check whether the constraints are satisfied or not.
based on the GA model and passes the new rates to SEAWAT.
Based on the SEAWAT results, new set of pumping rates are
3. Seawater intrusion model
formed, and SEAWAT is called again to compute the head dis-
tribution. This process is continued until an optimal solution SEAWAT-2000 (Langevin et al. 2003), which is based on dis-
is reached based on the objective function and the constraints. perse interface approach has been used in the present study, in
The time period to optimize is given by the user. The numerical order to simulate three dimensional, variable density, ground-
model will start running from the initial time period irrespec- water flow and transport. SEAWAT-2000 was designed by com-
tive of the time period to optimize and through this the numer- bining a modified version of MODFLOW-2000 (Harbaugh et al.
ical model will take care of the time relationship. SEAWAT uses 2000) and MT3DMS (Zheng and Wang 1999) into a single
ISH JOURNAL OF HYDRAULIC ENGINEERING 3
computer program. The governing equation for variable den- 3.1. Spatial and temporal discretization
sity flow in terms of freshwater head as used in SEAWAT is
A regularly spaced, finite-difference model grid was developed
(Guo and Langevin 2002).
to study the seawater–freshwater interface. The model grid is
[ ( 𝜌−𝜌 )] [ ( )] depicted in Figure 2. Each cell represents 100 × 100 m. The
𝜕hf 𝜕hf 𝜌 − 𝜌f 𝜕Z
𝜕
𝜌Kx +
f 𝜕Z
+
𝜕
𝜌Ky + grid consists of 92 rows and 53 columns. The top and bottom
𝜕x 𝜕x 𝜌f 𝜕x 𝜕y 𝜕y 𝜌f 𝜕y of the single layer model is based on borehole details from nine
boreholes. The top elevation varies from 0 to 9 m above MSL
[ ( )]
𝜕 𝜕hf 𝜌 − 𝜌f 𝜕Z 𝜕hf 𝜕𝜌 𝜕C
+ 𝜌Kz + = 𝜌Sf +𝜃 − 𝜌 s qs
(1) and bottom elevation varies from 1 to 16 m below MSL. Figure
𝜕z 𝜕z 𝜌f 𝜕z 𝜕z 𝜕C 𝜕t
3 shows the east–west section through Besant Nagar.
Temporal discretization was done by dividing the simula-
where x, y, z are orthogonal coordinate axes, aligned with
tion time into 72 monthly stress periods from January 2000 to
the principal directions of permeability; K is hydraulic con-
December 2005. The stress period is again divided into trans-
ductivity [LT−1]; S is specific storage [L−1]; t is time [T]; θ is
port steps. The transport step is automatically selected by the
effective porosity [dimensionless]; C is solute concentration
program to meet various stability constraints that are solution
[ML−3]; 〉s is fluid density source or sink water [ML−3]; 〉 is the
dependent.
density of the native aquifer water [ML−3]; 〉f is the density
of freshwater[ML−3], Z is the elevation at the measurement
point [L]. and qs is the volumetric flow rate of sources and 3.2. Initial and boundary conditions
sinks per unit volume of aquifer [T−1]. hf the equivalent fresh-
Boundary conditions were established to represent as closely as
water head.
possible the conceptual model of the flow system. The bound-
Downloaded by [La Trobe University] at 02:19 30 September 2017
i=1
ST
qit = f (ht−1
j ) From the simulation model (5)
∗
Figure 5. Location of control wells for calibration. 0 < qi,t < qi,t i = 1.................., n (7)
where Z(qt): the objective function for a given time period
t; htj: computed hydraulic head at the observation well j; h∗tj:
minimum allowed hydraulic head (MSL) at the observation
well j; qi,t: pumping rate from well i; qi∗ t: maximum allowed
pumping rate from well i; m: number of observation wells; n:
number of pumping wells.
The problem is formulated as an optimization problem to
maximize pumping considering the constraints on water levels
and pumping. The quantity of water being pumped from each
well during the year 2005 was calculated based on the given
rates and that was given as upper limit in the optimization
and the lower limit is set to zero. The constraint on hydraulic
head is given a safe value of 0.5 i.e. the hydraulic head in the
aquifer should be above 0.5 m of mean sea level so that there is
Figure 6. Comparison of observed and computed heads of observation well at no danger of saltwater intrusion. The primary objective of the
Besant Nagar. model is to maximize freshwater withdrawal to meet drinking
water needs from the aquifer, while maintaining hydraulic head
under permissible limit.
Application of genetic algorithm to constrained optimiza-
tion problems is often a challenging effort. There are several
approaches proposed in GA to handle constrained optimi-
zation problems. Among these methods, the most common
method in genetic algorithm to handle constraints is to use
penalty functions (Yeniay 2005). For the current study, the
optimization problem is transformed into an unconstrained
one by incorporating a penalty to the objective function. The
constraints are checked in the objective function evaluation
stage and the constraints for each management period are han-
dled by adding a penalty to the objective function if they are
not met. Penalty will be calculated based on the deviation of
the simulated head from the target head. Greater the violation,
Figure 7. The comparison of observed and computed TDS concentration of
observation well at Besant Nagar. higher will be the penalty assessed.
6 S. K. PRAMADA ET AL.
Downloaded by [La Trobe University] at 02:19 30 September 2017
Table 1. The initial and finally adopted parameter values. to prefer better solutions to worse ones. There are many selec-
tion procedures, including roulette wheel selection, stochastic
Sl.No Parameter Initial value Final calibrated value
universal selection, ranking selection, and tournament selec-
1 Hydraulic conductivity 65 50.2–65
(K in m/day) tion. These different selection procedures were tried for the
2 Specific yield 0.15 0.15 current problem. Stochastic selection was taking lot of time in
3 Effective porosity (θ) 0.2 0.2 terms of number of days to get the optimum, whereas roulette
4 Longitudinal dispersion 50 35
(αL in m) wheel selection was taking only 8 h. Thus, roulette wheel selec-
5 Transverse dispersion 5 3.5 tion was adopted in this selection process. The parent strings
(αT in m) of a string that will exist on the next generation are picked by
6 Density of freshwater(ρf 1000 1000
in kg/m3) a roulette wheel selection algorithm. This means that higher
7 Density of source(ρs in 1025 1025 fitness strings represent bigger parts of the wheel so they have
kg/m3) more chances to be picked.
8 Concentration of source 35,000 35,000
(Csmg/l)
4.5. Crossover operator
Table 2. Optimum pumping for each zone. After selection, individuals from the population are recom-
bined (or crossed over) to create new offspring. Recombining
Actual Pumping Optimal rate of
Number of during the Year pumping Year good individuals likely to create better individuals. Crossover
Zone wells 2005 (m3/day) (2006) m3/day exchanges portions of a pair of chromosomes at a randomly
Besant Nagar 4 8686.74 0 chosen point called the crossover point. The split position is
Thiruvanmiyur 2 2570.736 0 determined randomly. There are different crossover methods
Kottivakkam 3 1079.075 1070.6
Palavakkam 7 344.3517 317.8 available, namely single point, two point, n point, uniform,
Neelangarai 11 3570.467 3560.6 etc., Crossover does not occur always. Those chromosomes
Injambakkam 2 272.4218 270.3 with a higher fitness value are more likely to reproduce off-
spring. Crossover takes place based on a set probability, called
crossover probability. Two-point cross over is applied for the
4.4. Select solutions problem. Crossover probability was determined by conduct-
Selection allocates more copies of those solutions with higher ing sensitivity analysis. The crossover probability was varied
fitness values and thus imposes the survival of the fittest mech- from 0.6 to 1. For the current problem, crossover probability
anism of the candidate solutions. The main idea of selection is is fixed as 0.9.
ISH JOURNAL OF HYDRAULIC ENGINEERING 7