Beruflich Dokumente
Kultur Dokumente
155
INTRODUCTION
Groundwater extraction is required in many coastal areas where the freshwater supply
from surface sources is not adequate. However, excessive groundwater extraction may
lead to seawater intrusion into the aquifer, and hence excessive salinity. Therefore,
groundwater development in coastal aquifers must involve management strategies such
that the demand is met in terms of both quality and quantity. Groundwater manageOpen for discussion until 1 August 2004
156
S. V. N. Rao et al.
ment models for coastal aquifers should combine an accurate process model for
seawater intrusion with an appropriate technique for solving the optimization problem.
In this study, a nonlinear optimization model is developed as a tool to determine the
optimal operation strategy wherein the groundwater extraction is maximized, while the
seawater intrusion is minimized through a series of barrier wells.
The seawater intrusion phenomenon in coastal aquifers occurs as a result of reversal in
hydraulic gradients. There exists a transition zone between the freshwater and saltwater
zones, in which there is a gradual change in the density. Disperse interface or miscible
flow simulation models (Huyakorn et al., 1987; Anderson et al., 1988; Guo & Langevin,
2002) explicitly represent the dynamics of flow in this zone through an advection
dispersion equation. An alternative approach (Essaid, 1990) to the analysis of seawater
intrusion problems is based on the simplifying assumption that the transition zone can be
replaced by a sharp interface, when the width of the transition zone is small compared to
the thickness of the aquifer. The freshwater and saltwater flow domains are coupled
through an interfacial boundary condition of continuity of flux and pressure. The density
of the water in each zone is considered to be constant. Although both the approaches
account for saltwater dynamics, the miscible flow approach is closer to reality than the
sharp interface approach. However, use of the miscible flow approach in the management
models has been limited because of the high computational burden (Das & Datta, 1999).
The response of an aquifer system to specified recharge and pumping rates can be
determined using simulation models. However, coupled simulationoptimization (S/O)
models need to be used to arrive at management decisions during planning, design and
operation stages. Gorelick (1983), Peralta & Datta (1990), Hallaji & Yazicigil (1996),
Karatzas & Pinder (1996) and Zheng & Wang (2002), among others, have developed a
number of such groundwater management models. With regard to coastal aquifers in
particular, Willis & Finney (1988) developed an optimization model for minimizing
the seawater intrusion as well as the cost of pumping and recharge. In this model, the
sharp interface approach (Mercer et al., 1980) was used to simulate the seawater
intrusion. A reduced gradient quasi-Newton method was used for solving the optimization problem. Emch & Yeh (1998) developed a nonlinear multi-objective management
model for conjunctive use of surface water and groundwater in coastal areas. This
model was based on the SHARP interface approach for the aquifer response (Essaid,
1990), and the MINOS (Murtagh & Saunders 1980) nonlinear optimization package
for the solution of the optimization problem. Recently, Das & Datta (1999) developed
multi-objective management models involving groundwater development in a hypothetical coastal aquifer. These authors used an embedding approach and the MINOS
package for the solution of the optimization problem. The three-dimensional (3-D)
flow and transport equations were embedded as equality constraints in the optimization
model. The embedding procedure and the use of 3-D density-dependent flow equations
result in high computational burdens.
The S/O approach is appealing because it can readily use existing simulation
models and can account for the complex behaviour of a groundwater flow system in a
coastal aquifer. Although the S/O approach can identify the best management strategy
under a set of constraints, the computational burden is generally high and more so
when evolutionary optimization procedures are used. During the last 10 years, evolutionary or heuristic (non-exact) methods have become popular for solving the
nonlinear, non-convex optimization problems. In these cases, the presence of multiple
157
local minima poses problems if classical optimization techniques based on the gradient
search approach are used (Emch & Yeh 1998). Since no method exists that can
guarantee a global optimum, heuristic and evolutionary approaches are often referred
to as global optimization methods. Among these, the Simulated Annealing (SA) and
the Genetic Algorithms (GA) are the two most popular methods. Dougherty &
Marryott (1991) have demonstrated the application of SA to hypothetical groundwater
management problems in non-coastal regions. More recently, Zheng & Wang (2002)
have demonstrated the applicability of this approach to a field problem. These authors
used a response function to replace the simulator to reduce the computational burden.
In the present study, a coupled S/O approach is used to prioritize groundwater
development in a coastal aquifer system. A management problem is formulated with two
conflicting objectives involving maximization of groundwater development in coastal
zones while limiting the head and seawater intrusion to desired levels of salinity by the use
of barrier wells. Unlike many earlier models, the proposed model is based on a recently
developed 3-D density-dependent seawater intrusion model, SEAWAT (Guo & Langevin,
2002). Use of SEAWAT as the simulator is expected to result in a better representation of
the flow system, as compared to the sharp interface approach. Also, the SA technique is
used to solve the optimization problem in order to avoid problems common to classical
gradient search methods such as those used in MINOS. The coupled S/O model based on
the SA technique results in significant computational burden. This problem is tackled
herein by replacing the simulator with an artificial neural network (ANN), and the
computational burdens are kept within practical time frames for a desktop computer.
GOVERNING EQUATIONS
The 3-D density-dependent miscible flow and transport equations (Guo & Langevin,
2002) for a coastal aquifer may be written as follows:
()
t
Flow equation:
(q ) + q s =
Transport equation:
D(c) (vc) + q s =
(1)
c
t
(2)
k y P
k P
k P
, qy =
, q z = z + g ;
+
+ ; qx = x
x y z
y
z
x
is the variable fluid density [M L-3]; q is the specific discharge vector [L T-1]; is
the density of water entering from a source or leaving through a sink [M L-3]; qs is the
volumetric flow rate per unit volume of aquifer representing sources/sinks [T-1]; is
porosity [-]; t is time [T]; c is solute concentration [M L-3]; qx, qy, qz are the individual
components of specific discharge; is the dynamic viscosity [M L-1 T-1]; kx, ky, kz
represent intrinsic permeabilities [L2] in the three coordinate directions; g is the gravitational constant [L T2] and treated here as positive scalar quantity; P is the fluid pore
pressure [M L-1 T-2]; D is the dispersion coefficient [L2 T-1]; and v represents the
seepage velocity [L T-1].
158
S. V. N. Rao et al.
(3)
(a)
(b)
Fig. 1 Definition sketch of hypothetical 3-D aquifer system. (a) plan; (b) cross-section
AB.
159
1. Minimize pumping in extraction barrier wells near the coast along space and time:
N
J 1 = Qsin , j ,k
(4)
n =1 k =1 j =1 i =1
where Qsin , j ,k is the pumpage (decision variable) from the barrier wells located at
the node (i, j, k) during the time period n.
2. Maximize pumping in production wells at specified points:
N
J 2 = mQPn
(5)
n =1
where QPn is the pumpage from each well in a group of m production wells during
the time period n.
Subject to following constraints:
(a) Heads (h) at nodes along space and time should be greater than the specified value
hs. This constraint is imposed to limit the drawdown in the aquifer.
hin, j ,1 > hs
i, j, k and n
(6)
(b) The concentration (c) in production wells should be less than specified value cs.
cin, j ,k < c s
(7)
(c) Flow and transport equations should be satisfied. These are built into the
simulation model.
f (h, c, q) in, j ,k = 0
(8)
The methodology adapted in this study uses a coupled S/O approach. The SEAWAT
(Guo & Langevin, 2002) model is used for simulation. The responses of the SEAWAT
model over a range of pumping from the barrier and production wells are used for
training the ANN. The trained network behaves as the virtual simulator. The network
simulator is subsequently interfaced with the SA algorithm, which is used as the
optimizer. Pumping is treated as a discrete decision variable and, therefore, the model
is solved as a combinatorial nonlinear, non-convex problem. As two conflicting
objectives are involved, the second objective (maximizing pumpages from production
wells) is incorporated as a constraint, and a trade-off curve between the two objectives
is obtained. The simulator (SEAWAT model), the optimizer (SA algorithm) and the
artificial neural network (ANN) are briefly discussed below.
160
S. V. N. Rao et al.
SEAWAT model
The SEAWAT model (Guo & Langevin, 2002) was recently developed by combining
the popular MODFLOW (McDonald & Harbough, 1988) and MT3D (Zheng & Wang,
1998) models, with modifications to account for density variations between seawater
and freshwater. The pressure head is converted to equivalent freshwater head for the
variable water density in space and time. During any computational time step, the flow
field is first solved by MODFLOW and this is followed by the solution for concentration using MT3D. The updated density field is then calculated from the new concentrations and is incorporated back into MODFLOW as relative density difference terms.
The flow and transport equations are solved several times for the same time step
until the difference in fluid density between consecutive iterations is less than a userspecified tolerance.
SA algorithm
This is a heuristic algorithm in which each decision variable can take a discrete value
from a specified set of possible values. Each combination of decision variables is called
a configuration. The set of all possible configurations constitutes the configuration
space. The method first generates a random configuration (trial point) within the configuration space. Then a function call is made to the simulator to determine the state
variables. Values of the state variables are used to verify the constraints and to evaluate
the objective function. Further action depends on the following three possible scenarios:
1. Constraints are violated. In this case, the current trial point is rejected and a new point
is generated.
2. Constraints are not violated and the objective function value is better than the previous
best record. The trial point is accepted, the record for the best point is updated, and
then a new point is generated.
3. The trial point results in feasibility but the objective function value is worse than the
previous best value. In this case, the trial point is either accepted or rejected based on
the Metropolis criterion (Metropolis et al., 1953). For this purpose, a random deviate,
which is uniformly distributed in the interval (0, 1), is generated. The worse or the
uphill move is accepted if the above random deviate is smaller than the acceptance
probability. A new trial point is then generated. The probability for the acceptance of
an uphill move is equal to exp(C/T), where C is the difference in the objective
function values corresponding to the present and previous best configurations, and T is
a parameter called temperature. In the initial stages, a large percentage of downhill/
uphill moves are accepted by specifying a large value of T. The value of T is
progressively reduced as the trials proceed in order to reduce the acceptance probability. The acceptance probability of uphill moves eventually decreases to zero.
The aforementioned steps are repeated with the new trial point, and the process is
continued until equilibrium is achieved for the current value of T. It is assumed that the
equilibrium is achieved if the objective function value does not improve for a sufficient
number of iterations (chains). The temperature is then gradually reduced, and the entire
process is repeated for the new temperature. The iterative process is continued until the
termination criterion is met. The termination criterion is assumed to have been met if
the objective function value does not improve for four successive temperature reduc-
161
tions. More details are available elsewhere (Aarts & Korst, 1989; Dougherty &
Marryott, 1991; Teegavarapu & Simonovic, 2002).
The code for the SA algorithm developed here incorporates a perturbation procedure
called excursion limiting (Dougherty & Marryott, 1991) for generating a new random
configuration. In this procedure, the integer index level of each decision variable is
allowed to change randomly within l levels of its current value for a particular
configuration. The entire search space is covered in the initial stages by taking a large
value of l. The value of l is gradually reduced as the temperature is lowered, to reduce
the search space. The value of l tends to zero as the optimal (or near optimal)
configuration is reached.
ANN as the simulator
In this study, a large number of randomly generated configurations of the decision vector,
and corresponding responses of the SEAWAT model, are fed as inputs and outputs to train
the ANN. The trained ANN is then used instead of the SEAWAT model as a virtual simulator. An ANN attempts to mimic, in a very simplified way, the human mental and neural
structure and functions (Hsieh, 1993). It can be characterized as massively parallel
interconnections of simple neurons that function as a collective system (ASCE, 2000). The
network topology consists of a set of nodes (neurons) connected by links and usually
organized in a number of layers. Each node in a layer receives and processes weighted
input from the previous layer and transmits its output to nodes in the following layer
through links. Each link is assigned a weight, which is a numerical estimate of the connection strength. The weighted summation of inputs to a node is converted to an output
according to a transfer function (typically a sigmoid function). Most ANNs have three
layers or more: an input layer, which is used to present data to the network; an output
layer, which is used to produce an appropriate response to the given input; and one or
more intermediate layers, which are used to act as a collection of feature detectors. The
goal of the ANN here is to establish a relation of the form:
(Ym) = f(Xn)
(9)
n
m
where X is an n-dimensional input vector consisting of x1, x2, ..., xn; and Y is an
m-dimensional output or target vector consisting of resulting variables of interest y1, y2,
..., yn; and f() is the commonly used sigmoidal transfer function given by:
f(t) = 1/(1 + exp(t))
(10)
The network is trained generally, using a back propagation algorithm that will adjust
the weights and biases so as to minimize the error function given by:
E = ( yi ti )
P
(11)
where yi is the ANN output; ti is the desired output; p is the number of output nodes;
and P is the number of training patterns or data sets.
ILLUSTRATIVE APPLICATION OF THE MODEL
To illustrate the methodology, a hypothetical study area (Fig. 1) similar to that reported
by Das & Datta (1999) is used with finer spatial discretization. A seven-layer, 12-row
162
S. V. N. Rao et al.
and 22-column finite difference grid is constructed to represent a simplified 3-D aquifer
system. Six control wells (c1c6), two production wells (x7 and x8, pumping equal
discharge) and one existing production well (p9, with a fixed rate of pumping) are
considered. The sea face and all other boundaries are assumed vertical. The northsouth
and bottom boundaries are considered as no-flow boundaries. The left boundary is
assumed to be a time invariant constant head (3 m elevation) river boundary with zero
concentration. The sea face is assumed as a time invariant constant head boundary with
zero elevation and constant salinity of 35 kg m-3 representing seawater. The upper
phreatic surface receives a specified amount of recharge on a seasonal (six-monthly)
basis. Aquifer parameters similar to those in Das & Datta (1999) are adopted and are
listed in Table 1. Initially the aquifer geometry, boundary conditions, initial conditions
(starting heads and concentrations) and well locations are made symmetrical about one
axis as shown in Fig. 1. A pre-processor is used to create the required input files for
SEAWAT. The model is run using a false transient approach with only average
recharge for a long time period until steady state conditions in terms of heads and
concentrations are achieved.
Table 1 Parameters used in the SEAWAT model.
Parameters
Hydraulic conductivity in X, Y and Z directions
Porosity
Specific yield (unconfined aquifer)
Specific storage of confined aquifer
Molecular diffusion
Longitudinal and vertical dispersivity
Uniform rainfall recharge
Grid in X and Y directions
Grid in Z direction
Density difference ratio
Concentration of freshwater
Concentration of seawater
Density of seawater
Density of freshwater
Values
20, 20 and 0.2 m day-1
0.26
0.225
0.00044
7.7 10-5 m day-1
66.66 and 13.11 m
0.02 m per monsoon season
100 m
22.5 m
0.025
0
35 kg m-3
1025 kg m-3
1000 kg m-3
For any given set of pumpages, the SEAWAT model takes on average 4 s to execute
one stress period (180 days) involving solution of the flow and transport steps (the
MT3D model further divides each stress period into transport steps) on a desktop PC.
Since the optimization process by SA involves several thousands of function calls to the
simulator, a virtual ANN simulator model is developed as discussed before. For training,
two data sets are generated: one for the symmetrical aquifer system with only one stress
period and the second for the near-real aquifer system with three stress periods, both
beginning with the monsoon season.
To generate training sets for the ANN, pumpages are assigned randomly to all eight
locations (input variables: six control wells and two production wells as in Fig. 1). Each
decision variable is restricted to 10 possible discrete values (200, 400, 600, 800, 1000,
1200, 1400, 1600 and 1800 m3 day-1). In the first training set, the pumpage for the third
production well is set to zero, while a fixed value (1200 m3 day-1) is assigned to represent a
partially developed near-real aquifer system in the second training set. Under typical
163
Indian conditions, recharge occurs mostly during the monsoon season. Therefore, a
uniform recharge of 0.02 m per season was applied during the monsoon season (180 days).
No recharge is assumed during the non-monsoon season. No constraints are imposed for
generating data sets for the ANN training. The SEAWAT responses at points of interest for
the two cases are obtained by repeated execution of the model. The training sets included 8,
16 and 24 inputs (decision variables) corresponding to the first, second and third stress
periods. The output responses are heads at all nine pumping locations (first layer) and
concentrations at three production well locations (third layer) for each stress period.
For the two cases discussed previously, 4900 inputoutput data sets are generated.
The number of training sets (patterns) is generally a matter of concern when dealing with
observed data. In the present case this is not a problem as any number of data sets could
be generated through repeated execution of the SEAWAT model. The statistics of data
sets used for ANN training for the second case (near-real aquifer) are listed in Table 2.
The data sets cover the full range over which each input variable varied. This is because
ANNs are known to be good as interpolators rather than as extrapolators.
Table 2 Statistics of data sets from SEAWAT responses used for training ANN for the near-real aquifer system.
Well location
(see Fig. 1)
Head (m)
1
2
3
4
5
6
7
8
9
7
Max.
0.658
0.654
0.664
0.682
0.703
0.72
1.054
1.073
0.851
2.463
Max.
0.647
0.636
0.631
0.633
0.642
0.66
0.992
0.994
0.821
3.697
Max.
0.646
0.636
0.632
0.641
0.656
0.675
1.015
1.022
0.832
3.981
Min.
0.204
0.162
0.141
0.146
0.176
0.22
0.298
0.305
0.214
0.646
Mean
0.442
0.420
0.413
0.422
0.444
0.473
0.689
0.700
0.547
1.245
R2
0.995
0.992
0.999
0.991
0.992
0.992
0.996
0.989
0.988
0.978
Min.
0.171
0.137
0.126
0.139
0.167
0.21
0.278
0.292
0.187
0.537
Mean
0.404
0.382
0.374
0.384
0.407
0.437
0.637
0.648
0.501
1.396
R2
0.995
0.991
0.996
0.997
0.997
0.996
0.993
0.991
0.993
0.984
8
2.637 0.778
1.337
0.975
3.594 0.69
1.504
0.968
4.198
9
8.036 2.196
4.128
0.997
12.015 2.685 5.460
0.991
12.812
Max.: maximum value, Min.: minimum value, Mean: mean value in the data set (4900 patterns).
R2 refers to calibration of the ANN/ SEAWAT data set.
Min.
0.155
0.113
0.09
0.095
0.125
0.171
0.24
0.248
0.158
0.575
Mean
0.396
0.379
0.365
0.375
0.399
0.430
0.634
0.645
0.495
1.570
R2
0.995
0.991
0.986
0.991
0.993
0.992
0.995
0.981
0.991
0.984
0.722
3.224
1.699
6.390
0.976
0.991
The above data sets are standardized before ANN training is implemented. For this
purpose, the inputoutput data series (pattern) are scaled (xscaled) to values between zero
(0.0) and one (1.0) using the equation xscaled = (x xmin)/(xmax xmin), where x is the actual
value of the variable, and xmin and xmax are the minimum and the maximum values of the
variables respectively. This scaling is necessary because the sigmoid transfer function is
used in the network. A three-layer feed-forward network is trained using the ANN
toolbox of MATLAB (2000) to obtain the weights and biases of each network. The
network uses a sigmoidal transfer function (for six hidden neurons) and a linear function
(for one output neuron). The supervised training is accomplished with the help of a backpropagation algorithm (Marquardt-Levenberg) as implemented in MATLAB. Thus
several networks are trained. The training takes a few minutes to several hours depending on the number of inputs and outputs, i.e. the inputoutput vector size. For example,
for the third stress period, the training of heads with 24 inputs and six outputs takes
164
S. V. N. Rao et al.
R = 0.9800
R = 0.9834
0
3
4
3
The hypothetical symmetrical aquifer system considered here provides a basis to test the
management model developed using ANN/SA. The trained network is interfaced with the
SA code as discussed in the previous section. Only one stress period is considered and the
discharges from the production wells are set arbitrarily at 1800 m3 day-1 (within the range
of ANN training, i.e. 200, 400, 600, 800, 1000, 1200, 1400, 1600 and 1800 m3 day-1). The
constraint in respect of head at all the pumping locations is set at 0.2 m, and concentrations
in the two production wells are limited to 4 kg m-3. The annealing parameters are set by
trial as discussed in the next section. The management model ANN/SA is run to determine
the optimal configuration of pumpages in salinity control wells. Intuitively, since the
aquifer system is completely symmetrical about one axis in the middle in terms of
165
Table 3 Optimal pumpages, heads and concentrations in control wells for the symmetrical system with
constraint on all heads >0.2 m and salinity <4 kg m-3 in production wells.
Control wells (Fig. 1)
Pumpage (in m3 day-1)
Head (m)
Concentration (kg m-3)
0.573
13.217
2
200
0.550
15.811
3
600
0.534
19.328
4
600
0.534
19.328
5
200
0.550
15.811
0.573
13.217
boundary conditions, stresses, initial conditions and aquifer properties, the optimal solution
for any given equal pumpages in two production wells must also be symmetrical. The
results from the ANN/SA model, presented in Table 3, are consistent with this hypothesis.
Thus, the symmetrical aquifer system provides a basis to verify the conceptual model and
code although the solution is trivial. The optimal configuration is also verified for
constraints using the original SEAWAT model.
Near-real aquifer systems
Real aquifer systems are generally partially developed and are not symmetrical in terms
of boundary conditions, aquifer properties, layer type (confined/unconfined/semiconfined), stresses, and initial conditions in terms of heads and concentrations. In short,
they are complex. While some, if not all, of these can be incorporated into the
hypothetical aquifer system under consideration, only asymmetry in terms of stresses
(pumpages) is considered here to retain the simplicity in the analysis of the results. This
is achieved by assigning a fixed pumpage (1200 m3 day-1) to a third production well
located eccentrically but nearer to coast so as to induce seawater intrusion. The third
well is assumed to represent an existing well in a partially developed aquifer system, as
discussed previously. The ANN network from the second training set is interfaced with
SA to obtain the ANN/SA model for the near-real aquifer system. The constraints
pertaining to heads and concentrations for the near-real system are set moderately tight
at 0.2 m and 3.75 kg m-3 respectively at all locations, except at one location in terms of
concentration for the existing production well. The concentration level in the existing
production well (p9) is ignored in the present study. Nevertheless the salinity levels here
could also be constrained, as the same is included in the ANN training.
From the point of view of obtaining a trade-off curve for the multi-objective management model, the second objective of maximizing the pumpages from two production
wells (pumping at the same rate) is actually a constraint. Therefore, these pumpages
become known input values within the specified range for each individual run of the
single objective minimization problem. A typical optimal solution for a given pair of
pumpages in the two production wells (1600 m3 day-1) is shown in Table 4. For each
incremental value of the pumpages in production wells, an optimal configuration of
pumpages in the control wells for the three time periods is obtained. Thus a trade-off
curve limiting salinity levels to 3.75 kg m-3 is generated (Fig. 3). The first point on the
trade-off curve determines the maximum pumpage from the two production wells with
minimum pumpage from the barrier wells. Since the minimum possible pumpage from
the barrier wells can be zero, the solution is actually a global minimum. The remaining
points on the trade-off curve could be considered as near optimal solutions as the solutions cannot be verified. The trade-off curves could be used to prioritize groundwater
166
S. V. N. Rao et al.
Table 4 Optimal pumpages for near-real aquifer system in control wells, for production wells pumping
at 1400 m3 day-1 (constraint: heads >0.2 m and salinity <3.75kg m-3).
At the end of first stress
At the end of second stress
At the end of third stress
period (180 days)
period (360 days)
period (540 days)
Salinity
Pumping
Salinity
Pumping
Salinity
Pumping
(kg m3)
(m3 day-1)
(kg m3)
(m3 day-1)
(kg m3)
(m3 day-1)
c1
0.00
200.00
0.00
c2
0.00
0.00
0.00
c3
200.00
0.00
200.00
c4
200.00
0.00
0.00
c5
0.00
200.00
0.00
c6
200.00
0.00
0.00
x7*
1400.00
2.27
1400.00
3.22
1400.00
3.67
x8*
1400.00
2.19
1400.00
2.91
1400.00
3.44
p9 #
1200.00
7.73
1200.00
9.90
1200.00
11.49
Note: Barrier wells c1c6 indicate optimal pumpages.
* Production well pumpages (1400 m3 day-1) are fixed for each optimization run.
#
Existing well pumpages fixed throughout for all runs.
(volum e in m )
Location of
well
6.0x10
4.0x10
2.0x10
0.0
1.2x10
1.4x10
1.6x10
1.8x10
2.0x10
Fig. 3 Trade-off curve between total pumpages in production and control wells (in m3)
with concentration limited to 3.75 kg m-3.
development to meet demand in terms of quality and quantity. It is important to note that
the proposed model was formulated with equal pumpages from production wells to
ensure that a unique trade-off curve is evolved. Otherwise the solution would be a family
of trade-off curves. The effectiveness of barrier wells in controlling the salinity in production wells is demonstrated in Fig. 4, which shows the salinity level as a function of
the pumpage from the production wells for: (a) the case of no barrier wells, and (b) the
case of optimal pumping from the barrier wells. Typical isochlors (contours of equal
salinity) of 3.75 kg m-3 near the production well in the third layer at the end of 180, 360
and 540 days are shown in Fig. 5.
The annealing parameters are chosen based on the guidelines defined by
Dougherty & Marryott (1991), Press et al. (1996) and Cunha (1999). The initial
4.4
4.2
167
4.0
3.8
3.6
3.4
3.2
3.0
1200
1300
1400
1500
1600
3
1700
1800
-1
Fig. 5 Isochlors (3.75 kg m-3) at the end of 180, 360 and 540 days in the third layer
(plan view).
temperature (15 000) is set such that more than 80% of the configurations are accepted
in the beginning. Since the simulator is replaced with the ANN, longer chain lengths
are possible. The chain length (for equilibrium criterion) is set at 8090 times the
number of decision variables and the cooling factor (rate of reducing temperature)
equal to 0.4. The SA procedure is terminated when four successive temperature
reductions did not yield improvement in the solution. The evolution of the model
solution using the SA procedure is shown in Fig. 6.
168
2.0x10
1.5x10
1.0x10
5.0x10
S. V. N. Rao et al.
0.0
100
1000
10000
Number of evaluations
100
Percentage of acceptance
80
60
40
20
10000
1000
100
10
Temperature
COMPUTATIONAL TIME
The CPU (central processing unit) time depends on a number of factors. These include
the time consumed by the simulator, the number of decision variables, the tightness of
constraints, the speed of the processor, the efficiency of the perturbation procedure
(genetic rearrangement) and the annealing parameters (initial temperature, cooling
factor, number of configurations tested for each temperature, i.e. the chain length and
termination criterion) used. The SA procedure in the present methodology introduces a
computational time burden that has two distinct components.
The first component is due to the time consumed by the function calls to the
simulator and is associated with every trial configuration. This is virtually reduced to
near zero with the ANN as the simulator. Johnson & Rogers (2000) concluded that the
ANN replaces the full model. However, this is unlikely to be a rigorous statement.
This approach leads to a slightly altered feasible domain with the ANN as the
simulator, which may or may not contain the true global optimal solution.
169
The second component is the average time consumed for generating feasible configurations after satisfying all constraints until equilibrium and termination criteria are met.
It is kept to a minimum through efficient algorithmic guidance and through the perturbation procedure discussed above. The algorithmic guidance ensures that infeasible configurations are terminated at the earliest stage through efficient program coding.
The total CPU time is determined by the sum of the two components multiplied by
the total number of iterations or chains. The total number of iterations is problem
specific and, therefore, can be determined only after actual model execution.
In the present study, with the ANN as the simulator, the computational burden is
nominal. In all runs the CPU time did not exceed 10 min for moderately tight constraints
in respect of heads and concentrations for the trade-off curve shown in Fig. 3.
SUMMARY AND CONCLUSIONS
This study addresses the basic issue of seawater intrusion inherent to groundwater
development in coastal aquifers. The methodology involves control of seawater
intrusion through a series of barrier wells along the coast. An operational management
problem is formulated with two objectives involving maximization of groundwater
development through production wells, while minimizing the pumpage from the
barrier wells. Unlike many earlier models, the proposed model is based on a recently
developed 3-D density-dependent seawater intrusion model, SEAWAT. The SEAWAT
model is versatile and combines the existing codes of MODFLOW and MT3D that are
popular, well documented and easily accessible. It is demonstrated that ANN as a
universal approximator works very well as a substitute for numerical models to
manage computational burdens within practical time frames. Since ANNs are
essentially data driven models, their efficiency generally increases with the number of
data training sets. In the present study, as these are generated by repeated execution of
the SEAWAT model, any number of sets could be generated.
In the present study, the management model is cast as a combinatorial problem,
and the SA algorithm is used for solving it. Pumpages are considered as discrete
variables as pump capacities are discrete. The SA procedure presented herein can also
handle pumping as a continuous variable, with minor changes in the program.
However, the genetic rearrangement procedure adopted for perturbation is only
designed for discrete variables.
The utility of the study is demonstrated through the trade-off curve for prioritizing
groundwater development. Several trade-off curves for desired salinity levels in the
production well can be obtained. The trade-off curve is not linear. It is more expensive
to draw water at higher pumpages. In other words, some sacrifice has to be made
through barrier or cleanup wells to obtain cleaner water from production wells.
Acknowledgements The first two authors are grateful to Dr K. S. Ramasastri,
Director, National Institute of Hydrology (NIH), Roorkee, India, for the permission
and support to publish this work. All the authors are thankful to Sri P. R. Rao and Sri
D. M. Rangan of NIH, Kakinada for the neat figures and tables in the manuscript.
170
S. V. N. Rao et al.
REFERENCES
Aarts, E. & Korst, J. (1989) Simulated Annealing and Boltzmann Machines. John Wiley & Sons Inc., New York, USA.
Anderson, P. F., Mercer, J. W. & White, H. O. (1988) Numerical modeling of saltwater intrusion at Hallandale, Florida.
Ground Water 26(5), 619633.
ASCE (2000) Artificial Neural Networks in Hydrology. I: Preliminary concepts. II: Hydrologic applications. By the ASCE
Task Committee on Application of Artificial Neural Networks in Hydrology (Chairman: Prof. Rao S. Govindraju, Jr)
J. Hydrol. Engng ASCE 5(2), 115137.
Baxter, G. P. & Wallace, C. C. (1916) Changes in volume upon solution in water of halogen salts of alkali metals: IX. Am.
Chem. Soc. J. 38, 70104.
Cunha, M. D. C. (1999) On solving aquifer management problems with simulated annealing algorithms. J. Water Resour.
Manage. 13, 153169.
Das, A. & Datta, B. (1999) Development of multi objective management models for coastal aquifers. J. Water Resour.
Plan. Manage. ASCE 125(2), 7687.
Dougherty, D. E. & Marryott, R. A. (1991) Optimal groundwater management. 1. Simulated annealing. Water Resour. Res.
27(10), 24932508.
Emch, P. G. & Yeh, W. W. G. (1998) Management model for conjunctive use of coastal surface water and groundwater.
J. Water Resour. Plan. Res. ASCE 124(3), 129139.
Essaid, H. I. (1990) A multilayered sharp interface model of coupled freshwater and saltwater flow in coastal systems:
model development and application. Water Resour. Res. 26(7), 14311454.
Gorelick, S. M. (1983) A review of distributed parameter groundwater management modeling methods. Water Resour.
Res. 19(2), 305319.
Guo, W. & Langevin, C. D. (2002) User Guide to SEAWAT: a computer program for simulation of three-dimensional
variable-density groundwater flow. US Geol. Survey Open File Report 01-434.
Hallaji, K. & Yazicigil, H. (1996). Optimal management of a coastal aquifer in southern Turkey. J. Water Resour. Plan.
Manage. ASCE 122(4), 233244.
Hsieh, C. (1993) Some potential applications of artificial neural networks in financial management. J. Systems Manage.
44(4), 1215.
Huyakorn, P. S., Anderson, P. F., Mercer, J. W. & White, H. O. (1987) Saltwater intrusion in aquifers. Development and
testing of a three dimensional finite-element model. Water Resour. Res. 23(2), 293312.
Johnson, V. M. & Rogers, L. L. (2000) Accuracy of neural network approximators in simulation-optimization. J. Water
Resour. Plan. Manage. ASCE 126(2), 4856.
Karatzas, G. P. & Pinder, G. F. (1996) The solution of groundwater quality management problems with a nonconvex
feasible region using a cutting plane optimization technique. Water Resour. Res. 32(4), 10911100.
MATLAB (2000) Neural Network Tool Box for Use with MATLAB. User Guide. V4. The Mathwork Inc., Natick,
Massachussetts, USA.
McDonald, M. G. & Harbaugh, A. W. (1988) A modular 3D finite difference groundwater flow model. In: Techniques of
Water Resources Investigations of The United States Geological Society, book 6, Ch. A1. US Geol. Survey, USA.
Mercer, J. W., Larson, S. P. & Foust, C. R. (1980) Simulation of saltwater interface motion. Ground Water 18(4), 374
385.
Metropolis, N., Rosenbluth, A. W. & Teller, A. H. (1953) Equation of state calculations by fast computing machines.
J. Chem. Phys. 21(6), 10871092.
Murtagh, B. A. & Saunders, M. A. (1980) MINOS Manual. Technical Report 80-14, Syst. Optimization Lab., Dept. of
Operational. Res., Stanford, California, USA.
Peralta, R. C. & Datta, B. (1990) Reconnaissance-level alternative optimal groundwater strategies. J. Water Resour. Plan.
Manage. ASCE 116(5), 676692.
Press, W. H., Teukolsky, S. A., Vetterling, W. T. & Flannery, W. P. (1996) Numerical Recipes. Ch. 10, Simulated
annealing methods, 438439. Cambridge University Press, Cambridge, UK.
Teegavarapu, R. S. V. & Simonovic, S. P. (2002) Optimal operation of reservoir systems using simulated annealing. Water
Resour. Manage. 16, 401428.
Todd, D. K. (1980) Groundwater Hydrology. John Wiley & Sons, Inc., New York, USA.
Willis, R. & Finney, W. A. (1988) Planning model for optimal control of saltwater intrusion. J. Water Resour. Plan.
Manage. ASCE 114(2), 163177.
Zheng, C. & Wang, P. P. (1998) MT3DMS: A Modular 3-D Multispecies Transport Model for Simulation of Advection,
Dispersion and Chemical Reactions of Contaminants in Groundwater Systems. US Army Corps of Engineers,
Waterways Experiment Station, Vicksburg, Mississippi, USA.
Zheng, C. & Wang, P. P. (2002) A field demonstration of the simulation optimization approach for remediation system
design. Groundwater 40(3), 258265.