Professional Documents
Culture Documents
Bateni 2015
Bateni 2015
PII: S1568-4946(14)00666-8
DOI: http://dx.doi.org/doi:10.1016/j.asoc.2014.12.022
Reference: ASOC 2693
This is a PDF file of an unedited manuscript that has been accepted for publication.
As a service to our customers we are providing this early version of the manuscript.
The manuscript will undergo copyediting, typesetting, and review of the resulting proof
before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that
apply to the journal pertain.
Graphical abstract
t
ip
cr
us
an
M
d
p te
ce
Comparison of time-drawdown plots from graphical, NLP, and ACO techniques for sample confined (datasets C3),
unconfined (datasets U4), and leaky (dataset L1) aquifers.
Ac
1
Page 1 of 33
Highlights
• We proposed an Ant Colony Optimization (ACO)to estimate the aquifer hydraulic porperties;
• ACO is efficient, and reliable for estimating various aquifer hydraulic parameters from the
time‐drawdown data;
• ACO provides better estimation of aquifer properties than the graphical and NLP techniques.
t
ip
cr
us
an
M
d
p te
ce
Ac
2
Page 2 of 33
Evaluation of Methods for Estimating Aquifer Hydraulic Parameters
1
Department of Civil and Environmental Engineering, University of California Los Angeles, Los
t
Angeles, CA, USA
ip
2
Department of Civil, Surveying and Environmental Engineering, University of Newcastle, Callaghan,
cr
NSW, Australia
3
Department of Civil Engineering, Sharif University of Technology, Tehran, Iran
us
4
Griffith School of Engineering, Griffith University Gold Coast Campus, Queensland, QLD 4222,
Australia
5
NOAA-Cooperative Remote Sensing Science & Technology Center (NOAA-CREST), City
an
University of New York, NY, USA
by graphical methods that are approximate and time-consuming. As a result, nonlinear programming
(NLP) techniques have been used extensively to estimate them. Despite the outperformance of NLP
te
approaches over graphical methods, they tend to converge to local minima and typically suffer from a
convergence problem. In this study, Genetic Algorithm (GA) and Ant Colony Optimization (ACO)
p
methods are used to identify hydraulic parameters (i.e., storage coefficient, hydraulic conductivity,
ce
transmissivity, specific yield, and leakage factor) of three types of aquifers namely, confined,
unconfined, and leaky from real time-drawdown pumping test data. The performance of GA and ACO
is also compared with that of graphical and NLP techniques. The results show that both GA and ACO
Ac
are efficient, robust, and reliable for estimating various aquifer hydraulic parameters from the time-
drawdown data and perform better than the graphical and NLP techniques. The outcomes also indicate
that the accuracy of GA and ACO is comparable. Comparing the running time of various utilized
methods illustrates that ACO converges to the optimal solution faster than other techniques, while the
graphical method has the highest running time.
Keywords: Aquifer Hydraulic Parameters, Ant Colony Optimization (ACO), Genetic Algorithm
(GA), Nonlinear Programming (NLP), Pumping Test.
3
Page 3 of 33
1. Introduction
In many parts of the world, groundwater has been used as a reliable source of water for numerous
purposes such as irrigation, and domestic and industrial uses. Therefore, it is necessary to
appropriately model the complex groundwater system in order to properly manage this vital resource.
t
To do so, an accurate specification of aquifer hydraulic parameters such as transmissivity (T),
ip
hydraulic conductivity (K), storage coefficient (S), leakage factor (B), and specific yield (Sy) is needed
cr
since these parameters are commonly required in groundwater/aquifer flow modeling [1-7]. These
parameters can be measured by laboratory experiments or in-situ tests; however, the former is
us
inaccurate, and the latter is expensive.
An alternative option for the estimation of the aquifer parameters is through inverse modeling. In
an
this procedure, some variables that can be obtained more accurately (e.g., hydraulic head) are
measured; then, the aquifer parameters are inversely estimated from the measured variables. Pumping
M
test is the most commonly used and standard technique for inverse modeling of groundwater
parameters. In this test, time-drawdown measurements are analyzed via the analytical solutions such
d
as Theis, corrected Theis, and Hantush models depending on the aquifer type in which the test is
te
performed [8].
p
Several approaches have been proposed to solve the above-mentioned inverse problem and
ce
obtain aquifer hydraulic parameters. The first group of approaches is graphical. The aquifer hydraulic
parameters estimated by the graphical technique are of questionable reliability since a perfect curve
Ac
matching is not anticipated and the match is typically approximate [1]. The traditional graphical
technique is time-consuming, burdensome, and subjective [8]. The second group of approaches
attempts to estimate aquifer parameters using nonlinear programming (NLP) techniques such as the
steepest descent method, conjugate gradient method, Gauss–Newton method, Marquardt algorithm,
etc. Although NLP methods often outperform the graphical approaches, they suffer from a number of
shortcomings. Aquifer models are typically non-convex and nonlinear, and for such models, classical
gradient-based optimization techniques may result in local optimum values rather than global ones.
4
Page 4 of 33
The gradient based techniques may cause instability and convergence problems due to the numerical
difficulties. Moreover, their convergence to a global optima highly depends on the selection of initial
search points. Furthermore, these techniques cannot find the local gradient for the problems with
t
ip
To overcome the aforementioned shortcomings, the non-traditional methods such as
evolutionary algorithms can be applied as they can handle highly nonlinear, even non-smooth
cr
problems and converge to the global rather than local optimum. In this study, two well-known
us
evolutionary algorithms namely, Genetic Algorithm (GA) and Ant Colony Optimization (ACO) are
proposed to estimate the hydraulic parameters of three major types of aquifers (i.e., confined,
an
unconfined, and leaky). An objective function is formulated for the inverse estimation using the
observed hydraulic heads from pumping tests. Thereafter, GA and ACO are utilized to minimize the
M
objective function by tuning the aquifer hydraulic parameters. These techniques have several
advantages over gradient-based methods [10, 11]: (1) They do not require a continuous objective
d
function; (2) They often can find a near global optima and do not get trapped in a local-optimal
te
solution; (3) They provide a number of near global optima solutions and hence give users the ability to
GA has been applied in a number of groundwater optimization problems. El Harrouni et al. [12]
ce
used GA to manage pumping from wells and to estimate hydraulic parameters in heterogeneous
aquifers. In a similar effort, Lingireddy [13] integrated GA and neural network to obtain aquifer
Ac
hydraulic parameters. Samuel and Jha [8] and Jha et al. [4] examined effectiveness of GA in
estimating aquifer parameters. They showed that GA is a viable and efficient technique that often
overcomes the subjectivity, high computational burden and ill-posedness of traditional optimization
methods. In a more recent attempt, Rajesh et al. [14] utilized GA, graphical, and sequential
5
Page 5 of 33
ACO has been successfully used in water resources optimization problems such as minimizing
the capital costs associated with water distribution systems [15-18], design of irrigation networks [19],
optimal groundwater monitoring design [20], multi-purpose reservoir operation [21, 22], optimal
design of open channels [23], underground seepage flow [24], and saltwater intrusion in the coastal
t
aquifers [25]. To the best knowledge of the authors, the only study on the estimation of aquifer
ip
parameters with ACO was conducted by Li et al. [26], wherein the hydraulic conductivity and storage
cr
coefficient of a confined aquifer were estimated in a synthetic experimental framework. In comparison
us
to their study, the present study uses ACO to estimate the hydraulic parameters of three aquifer
systems (confined, unconfined and leaky) from real-world time-drawdown pumping test data.
an
Moreover, the results of ACO are compared with those of graphical and NLP techniques. Each
method is applied to 15 sets of real time-drawdown pumping test data (5 sets of data for each aquifer
M
type). These data sets are taken from [27-33].
This paper is organized as follows. In section 2, the analytical models of groundwater used in
d
this study are reviewed. The graphical, NLP, GA, and ACO methods are outlined in section 3. Section
te
4 presents the formulation of the objective function. Section 5 provides the results and discussions.
Pumping test is a well-known tool to characterize a system of aquifers. The test is conducted by
pumping water from one well and observing the response of aquifer (drawdown) in observation wells.
Ac
The hydraulic-head drawdowns from the pumping test can be used for aquifer characterization and the
estimate of parameters such as transmissivity (T), storage coefficient (S), hydraulic conductivity (K),
leakage factor (B) and specific yield (Sy). Herein, the theoretical models for response of idealized
The well-known Theis [34] solution describes radial flow to a fully penetrated pumping well in
a homogeneous, isotropic confined aquifer of infinite horizontal extent which is pumped at a constant
6
Page 6 of 33
rate Q. Theis solution that is considered as one of the fundamental breakthroughs in the development
Q
s( r, t ) = W (U ) (1)
4π T
t
ip
where s(r, t) is the drawdown at a radial distance r from the pumping well at time t since the beginning
of pumping, T is the transmissivity, S is the storage coefficient, U = r 2 S / 4Tt , and W(U) is the Theis
cr
well function, which is given by
us
∞ e− x
W (U ) = ∫ dx (2)
U x
an
In this study, the Theis well function W(U) is calculated via the Taylor series expansion for small
values of U ( 0 < U ≤ 1 ) and the Gauss-Laguerre quadrature for large values of U ( U > 1 ) [35].
M
Unconfined aquifers have a different response to the pumping test compared to confined
aquifers. Therefore, the Theis model should be corrected in order to be used for unconfined aquifers.
d
The corrected Theis model replaces s(r, t) on the left hand side of Eq. 2 with s ′(r , t ) , where s ′ is
te
( )
equal to s − s 2D0 , and D0 is the initial saturated thickness of the unconfined aquifer [36, 37].
2
p
Theis and corrected Theis models are used in this study for confined and unconfined case studies,
ce
respectively.
Ac
The leaky-aquifer theory has been evolved from the Theis equation. In two sets of papers,
Hantush and Jacob [38] and later Hantush [39, 40], the original differentiation between the Theis
response for confined aquifers and the one for leaky aquifers was discussed. The analytical solution of
Hantush and Jacob [38] can be couched in the same form as the Theis solution (Eq. 1) but with a more
complicated well function. In analogy with Eq. (1), the drawdown is computed via,
7
Page 7 of 33
Q
s( r, t ) = ( ) W (U , r / B ) (3)
4π T
where B is the leakage factor and W (U , r / B ) is known as the leaky well function, which is given by
∞
W (U , r / B) = ∫ (1 / y ) exp(− y − β 2 / y )dy
t
(4)
ip
U
cr
in which β = r / (2 B ) .
us
3. Methods
an
3.1. Graphical Method
As stated in the introduction, the first group of approaches for retrieving aquifer parameters is
M
graphical. In order to determine aquifer parameters such as transmissivity (T) and storage coefficient
(S), the pumping test is conducted and the drawdowns in a pumped well and in several observation
d
wells are recorded at suitable time intervals. The recorded time-drawdown data are analyzed by the
te
standard graphical methods designed for confined, unconfined and confined aquifers [27, 29, 31].
These graphical methods basically involve matching of Type curves (which vary from the type of
p
aquifers) with the corresponding field curves, finding out a match point and then calculating aquifer
ce
parameters using the Theis model (for confined and unconfined aquifers) or Hantush and Jacob model
(for leaky aquifers). In this study, the time-drawdown pumping test data of confined, unconfined and
Ac
leaky aquifers were graphically analyzed using the commonly used AquiferTest software developed
by Waterloo Hydrogeologic Incorporation (WHI), Canada. It is a graphical oriented package, and has
a wide range of methods for analyzing pumping test data of a variety of aquifers. For detailed
(http://www.swstechnology.com/groundwater-software/pumping-test-analysis/aquifertest-pro).
8
Page 8 of 33
3.2. Nonlinear Programming (NLP)
As mentioned in the introduction, the second group of methods for estimating aquifer
parameters is based on nonlinear programming techniques. In this study, the well-known GAMS
(General Algebraic Modeling System; [41]) software is employed to estimate aquifer hydraulic
t
parameters from pumping test data using its NLP solvers. GAMS is a mathematical programming and
ip
optimization modeling system that has been used widely in the water resources and agricultural
cr
economics research communities [42-47].
us
GAMS has several built-in solvers for solving NLP problems, and the program itself tries to use
the best solver for each problem. One of the well-known NLP solvers in GAMS is CONOPT. Our
an
NLP problems in GAMS are solved using CONOPT. This solver is mainly developed for models with
smooth functions, but it can also be used for models with a non-differentiable function. CONOPT3,
M
the most recent version of the solver, is a gradient based algorithm that executes a sequential quadratic
programming (SQP) component. The SQP module utilizes the second derivative to obtain better
d
search directions [48]. The hydraulic parameters in our case study problems are found via a GAMS
te
optimization code. Finally, the results are compared with those of the other two methods.
which have highly adaptive behaviors will have relatively a larger number of offspring and
determines the number of chromosomes in each population. In GA, parameters are presented in form
of chromosomes which, in turn, consist of genes. Traditionally, GA has been developed by applying
binary coding in which a chromosome is represented by a string of binary bits, e. g., [00101001], that
9
Page 9 of 33
A simple flowchart for GA is shown in Figure 1. The first step in GA is to initialize population
randomly. Indeed, all chromosomes in the first population are determined randomly. Then, for each
chromosome, parameters are determined by decoding genes into real values. In the next step,
chromosomes are evaluated by calculating the value of objective function for each of them.
t
GA operators mimic natural processes. The main GA operators are selection, crossover and
ip
mutation. Selection is the process by which chromosomes are chosen for participation in the
cr
reproducing process. By applying selection operator, two chromosomes usually called parents, are
us
selected based on their objective function values. Crossover involves a pair of chromosomes
exchanging a portion of their bit sequence at randomly set cutting positions. The idea behind crossover
an
is that, by exchanging blocks of bits between two chromosomes that performed well, the GA attempts
to produce two new chromosomes that preserve best materials from two parents. There are a few
M
crossover methods including one-point crossover, two-point crossover and uniform cross over [49].
The main difference between these methods is number of cutting positions. For instance, one-point
d
crossover has one cutting position while two-point crossover has two cutting positions. Figure 2 shows
te
schematically the one-point cross over. Chromosome 1 and 2 are selected as parents and then a cutting
position is selected randomly (Fig. 2). Finally, the genes on the right hand side of the position are
p
exchanged.
ce
genes. Crossover and mutation operations do not allow GA to get trapped in local optima and search
the decision space more vigorously and efficiently. As it is shown in Figure 1, selection, evaluation,
crossover and mutation repeat until a termination criterion is met. Termination criterion is usually
defined as reaching the maximum number of generations or if there is no changes in the value of
Following Samuel and Jha [8] and Jha et al. [4], crossover and mutation probabilities of 0.8 and
0.015 are chosen. Population size and number of generations are set to 80 and 100, respectively [4, 8].
10
Page 10 of 33
The genetic algorithm is terminated when the maximum number of generations (i.e., 100) is reached.
The readers are referred to Goldberg [49] and Michalewicz [50] for a full description of GA.
[Figure 1 here]
t
ip
[Figure 2 here]
cr
3.4. Ant Colony Optimization (ACO)
us
Ant colony optimization (ACO) is a recently developed heuristic optimization method. It was
inspired by the fact that some kinds of ants are blind but they can find the minimum path between
an
their nest and food. This occurs because of chemical substances called pheromones that ants deposit
method called Ant System (AS) to solve the travelling salesman problem (TSP) and job-shop
d
scheduling problem (JSP). The results showed that ACO dominated all other classical or heuristic
te
methods solutions [52]. Since then ACO has been applied in a variety of engineering applications [16,
p
Abbaspour et al. [53] developed a new approach based on the ACO concept and successfully
tested the feasibility of their approach by estimating unsaturated soil hydraulic parameters. The main
Ac
difference of Abbaspour et al. [53] method with the traditional ACO is that the initial range of
parameters in their technique can be adjusted in the optimization process. For instance, the end of
parameters intervals may be eliminated or expanded. This feature makes their method more flexible to
deal with the model parameters range identification and to more robustly explore promising search
spaces by narrowing down the search space. In this study, the Abbaspour et al. [53] ACO algorithm is
used due to its distinct advantage compared to the traditional ACO method.
11
Page 11 of 33
The first step of ACO is to represent the search space as a graph. For this reason, each parameter
is split into a specific number of segments with a representative value assigned to each segment. Ants
travel between nodes corresponding to different parameters to define a route as illustrated in Figure 3.
In this figure, each of five parameters, V1 to V5, are discretized into six segments. In addition, three
t
sample routes are presented which show how each ant assigns a value to a parameter.
ip
[Figure 3 here]
cr
When ants travel on routes, the objective function of each route is calculated based on its
us
associated parameters. Then ants deposit pheromone on their routes according to the obtained
objective function value in a way that routes which resulted in larger objective function values receive
an
less pheromone. The following equation is used to update pheromone on segments
⎧ gu − T
M
⎪exp(4.6( )) gu ≤ T
τ u (I ) = ⎨ g min − T (5)
⎪0 gu > T
⎩
d
te
corresponding objective function value of route u, g min is the minimum value of the objective
p
ce
function obtained in the current iteration and T is a critical value of the objective function, above
σg
T = g min + CT (6)
μg
where CT is a constant which is determined to be 0.5, while σ g and μ g are the mean and standard
The main step in the algorithm is to update parameters’ intervals based on the amount of
pheromones on each segment. The idea behind this step is to use scoring strategy to eliminate or
expand segments. Once the scoring is completed, the segments which have small scores will be
12
Page 12 of 33
deleted from both ends of the intervals. On the other hand, if segments on either end of parameter
intervals have a higher score, the parameter could be extended in that direction, i.e., add specific
A score is calculated for each segment of a parameter by the following expression. Note that any
t
single segment may be the crossroad of many ant routes.
ip
(φij ) A (σ ij ) N
Sij = (7)
cr
∑∑
i j
(φij ) A (σ ij ) N
us
where Sij is a score, σ ij is the standard deviation of pheromones, φ ij is accumulated pheromones on
N = CN
σ ij
μij an (8)
M
where CN is a constant set equal to 0.3. In this study, number of ants is set to 100 and each parameter
d
is discretized into 500 segments.
te
4. Objective Function
p
The objective function g is defined as the sum of squared difference between the observed and
ce
calculated drawdowns:
Ac
n
g = Min( ∑ ⎡⎣ so ( t ) − sc ( t ) ⎤⎦ )
2
(9)
t =1
computed drawdown from a suitable model (Theis, Corrected Theis or Hantush Model) for the given
aquifer type, and n is the total number of observed drawdown data. The aim is to minimize the
objective function through estimating the optimum values for hydraulic parameters of aquifers. Figure
shows the flowchart of ACO algorithm for finding the hydraulic parameters of an aquifer. This
13
Page 13 of 33
flowchart is based on the ACO algorithm described in section 3.3. The iteration in the flowchart
continues until a termination criterion is satisfied. The algorithm terminated when the criterion
g = 10−4 m 2 is fulfilled or the maximum number of iterations reaches a prescribed value of 100. The
variables Q, r, t, and So (shown in Figure 2) are inputs for confined and leaky aquifers. For unconfined
t
ip
aquifers, in addition to the abovementioned variables, D0 is required as input.
[Figure 4 here]
cr
The performance of graphical, NLP, GA, and ACO methods is evaluated using the mean
us
absolute error (MAE) and the root mean square error (RMSE) metrics. These metrics are used to
assess the drawdown estimates from each approach. The coefficient of determination (R2) of linear
an
regression line between the estimated drawdown from the graphical, NLP, GA, and ACO models and
the observations is also used as a measure of performance. The low MAE and RMSE as well as high
M
R2 show that the estimated hydraulic parameters are close to their real physical values.
The AquiferTest software along with the NLP, GA, and ACO computer programs were run to
te
optimize aquifer parameters for five time-drawdown datasets associated with each type of aquifer
p
systems (confined, unconfined, and leaky). Thus, we analyzed a total of fifteen sets of time-drawdown
ce
pumping test data for three aquifer systems using graphical, GA, ACO, and NLP approaches. In this
study, the pumping test datasets associated with confined, unconfined, and leaky aquifers are denoted
Ac
respectively by C1-C5, U1-U5, and L1-L5. The datasets C1, C4, U1, U2, L1, and L2 are taken from
unpublished sources, C2 and C3 are obtained from Batu [27] and C5 is from Todd [31]. The datasets
U3, U4, U5, L3, L4, and L5 are obtained respectively from Walton [33], Fetter [29], Waltson [32],
Raghunath [30], Charbeneau [28], and Fetter [29]. These datasets in varying hydrogeologic conditions
allow us to compare the applicability, competence and robustness of the utilized methods in estimating
the hydraulic parameters of different types of aquifers. The statistical performance metrics mentioned
14
Page 14 of 33
in the previous section are used for this purpose. Finally, time-drawdown estimates from the graphical,
The datasets C2 and C3 are obtained from Batu [27], C5 datset is taken from Todd [31] and C1
t
ip
and C4 from unpublished sources. The estimated storage coefficient and transmissivity of five
explored confined aquifers from the graphical, NLP, GA, and ACO methods are shown in Table 1.
cr
These hydraulic parameters are used to compute drawdowns in the observation well at different times
us
via the Theis model (equation 1) and compare them with corresponding measurements. Statistical
indices (i.e., MAE, RMSE, and R2) associated with the computed drawdowns in the observation well
an
are presented in Table 1.
The storage coefficient and transmissivity estimates from the ACO vary respectively from
M
1.56 × 10 −4 to 5.30 × 10 −3 and 236 to 5559 m2/day. These values are in good agreement with those
obtained by the graphical, NLP, and GA methods. However, the ACO and GA results for all confined
d
aquifers datasets (C1-C5) are better than those of the graphical and NLP approaches (Table 1). The
te
MAE and RMSE of ACO and GA are always less than the ones from the graphical and NLP methods,
p
implying that ACO and GA can retrieve the hydraulic parameters of confined aquifers more
ce
accurately. As shown in Table 1, for C1, C2, C3, C4, and C5 datasets, the computed drawdown from
ACO (GA) has a RMSE of 0.2737 (0.2561), 0.0445 (0.0454), 0.0398 (0.0378), 0.0230 (0.0257), and
Ac
0.0051 m (0.0052 m), which is a 14% (20%), 34% (33%), 19% (23%), 56% (51%), and 41% (40%)
reduction of the graphical RMSE of 0.3184, 0.0676, 0.0490, 0.0521, 0.0086 m. Analogously, for the
abovementioned datasets, the MAE of estimated drawdown decreases from 0.2441, 0.0586, 0.0441,
0.0338, and 0.0067 m to 0.2372 (0.2216), 0.0388 (0.0395), 0.0321 (0.0315), 0.0163 (0.0193), and
0.0030 m (0.0033 m) by using ACO (GA) rather than the graphical approach. As can be seen in Table
1, the results of ACO and GA are close and comparable. In some of the confined aquifers (C2, C4, and
C5) ACO outperforms GA, and in others (C1 and C3) GA performs better. The ACO and GA
15
Page 15 of 33
methods have also the highest R2 value for all of confined aquifers datasets, showing that the
drawdown estimates from GA and ACO are more correlated to the observations than the ones from the
graphical and NLP techniques. Especially, for the C4 aquifer, the estimated storage coefficient and
transmissivity from ACO are significantly different from those of the other two methods. Given the
t
much better performance of ACO method (based on the statistical indices for the C4 aquifer), its
ip
hydraulic parameters estimates are more reliable and the NLP method has likely found a local
cr
minimum of objective function.
us
Overall, the statistical metrics suggest that ACO and GA are more efficient techniques for the
retrieval of hydraulic parameters of confined aquifers. Another conclusion from Table 1 is that the
an
errors associated with the NLP approach are less than those of graphical method and thus NLP can
estimate the storage coefficient and transmissivity more reliably than the graphical method.
M
[Table 1 here]
d
Table 2 lists the estimated specific yield and hydraulic conductivity values of the five
investigated unconfined aquifers from the graphical, NLP, GA, and ACO methods. The performance
p
metrics of each technique are also presented in Table 2. Based on the ACO results, the specific yield
ce
of the unconfined aquifers ranges from 3.38 × 10 −4 to 6.93 × 10 −2 , and the hydraulic conductivity
ranges from 23 to 169 m/day. The retrieved specific yield and hydraulic conductivity from graphical,
Ac
The ACO and GA methods perform better than the other two techniques. For all unconfined
aquifers, the MAE and RMSE values of ACO and GA are less than the corresponding values of the
other two methods. For U1, U2, U3, U4, and U5 datasets, the RMSE of estimated drawdown from
ACO (GA) is respectively 0.0194 (0.0300), 0.0801 (0.0671), 0.0314 (0.0320), 0.0265 (0.0403), 0.0103
m (0.0103 m), which is a factor of 67 (50), 53 (61), 37 (30), 65 (47), and 48 (50) less than the
corresponding RMSE estimates from graphical approach and a factor of 40 (35), 53 (60), 3 (1), 43
16
Page 16 of 33
(15), and 2 (5) less than the corresponding RMSE estimates from NLP method. These results prove
the capability of the ACO and GA methods in finding more accurate estimates of specific yield and
hydraulic conductivity of unconfined aquifers. Compared to GA, ACO gives slightly more accurate
aquifer parameters estimates for four of the unconfined aquifers (U1, U3, U4, and U5 datasets). While,
t
in one of the unconfined aquifers (U2 dataset), GA generates better results.
ip
Similar to the confined aquifer datasets, the NLP outperforms the graphical technique. This
cr
finding is consistent with that of Neuman et al. [56]. Aquifers are typically heterogeneous media and
us
thus their properties vary spatially such that cannot be described with certainty. Neuman et al. [56]
confirmed that time-drawdown data from randomly heterogeneous aquifers are hard to interpret
an
graphically.
[Table 2 here]
M
5.3. Leaky Aquifers
d
The storage coefficient, transmissivity, and leakage factor estimates from graphical, NLP, GA
te
and ACO approaches for the leaky aquifers are indicated in Table 3. Based on ACO results, the
storage coefficient varies from 2.25 × 10 −5 to 1.05 × 10 −3 , the transmissivity from 25.2 to 960.4
p
m2/day and leakage factor from 221.0 to 9002.6 m. The MAE and RMSE of ACO drawdown
ce
estimates are compared to those from the graphical, NLP, and GA methods (Table 3). In analogy to
the results of the confined and unconfined aquifers (Tables 1 and 2), ACO and GA perform better than
Ac
the graphical and NLP techniques. The errors of ACO and GA are always less than those of the other
two techniques. For L1, L2, L3, L4, and L5 aquifers, the ACO (GA) approach reduces the RMSE of
drawdown estimates from the NLP technique by 36% (47%), 27% (9%), 41% (12%), 39% (16%), and
55% (63%) and from the graphical method by 47% (55%), 35% (18%), 40% (12%), 64% (50%), and
58% (66%), respectively (Table 3). Similar drop is also evident in the MAE associated with ACO and
GA. Compared to the graphical and NLP methods, the ACO and GA estimates of drawdown have also
higher values of R2. These outcomes demonstrate the graphical and NLP techniques are not as
17
Page 17 of 33
efficient as ACO and GA for estimating hydraulic parameters of leaky aquifers. Likewise, Trinchero
et al. [57] showed that when various methods were applied to heterogeneous media of leaky aquifers,
each method could provide different estimates of hydraulic parameters. Comparing results of ACO
and GA reveals that none of them is superior to the other one. For L1 and L4 aquifers, GA slightly
t
outperforms ACO, but for L2, L3, and L4 aquifers, ant colony optimization method gives better
ip
results.
cr
[Table 3 here]
us
5.4. Visual Examination of Retrieved Time-Drawdown Curves
an
The efficiency of ACO and NLP optimization techniques as well as the graphical method in
estimating the hydraulic parameters of confined, unconfined and leaky aquifers is also assessed by
M
visual comparison of retrieved time-drawdown curves from these three techniques with observations.
The comparisons are illustrated in Figure 5. Since the time-drawdown curves from GA and ACO are
very close, only the one obtained by ACO is shown herein. As can be seen in these plots, drawdown
d
retrievals from ACO have always the best agreement and correlation with the observations. In all the
te
cases, the second best solution is obtained from the NLP method and the last one from the graphical
p
method. These results are consistent with the outcomes of the comparisons based on the statistical
ce
[Figure 5 here]
Ac
Besides the accuracy of graphical, NLP, GA, and ACO methods in estimating aquifer hydraulic
parameters, their running time is an important issue in groundwater engineering problems. It is of our
interest to understand which of the explored methods runs faster, especially when there are a large
The developed computer programs were executed on the WINDOWS operating system (Xeon
5128, 256 GB RAM, FSB 1066 MHz, 40 GB HDD), and their running times for different aquifer
18
Page 18 of 33
types are shown in Table 4. Each execution time for confined, unconfined, and leaky aquifers
represents the mean of running times for C1-C5, U1-U5, and L1-L5 datasets, respectively. As
illustrated, the ACO approach converges to the optimal solution faster than other methods, while the
graphical technique is the slowest one. On overage, for the five explored confined aquifers, the
t
running time of ACO is 57%, 26%, 37% less than that of graphical, NLP, and GA methods.
ip
Corresponding values are 53%, 20%, and 33% for the unconfined aquifers and 63%, 33%, and 42%
cr
for the leaky aquifers.
us
6. Conclusions
an
In this study, the efficiency of ant colony optimization (ACO), which usually ensures near-
based on the misfit between drawdown observations from pumping test and corresponding drawdown
d
estimates from Theis, corrected Theis, or Hantush models, depending on the aquifer type. The ACO
te
technique is utilized to estimate the hydraulic parameters of various aquifer systems by minimizing the
objective function. Performance of ACO is compared with that of genetic algorithm (GA), graphical
p
and nonlinear programming (NLP) methods using fifteen sets of published and unpublished real time-
ce
drawdown data of three aquifer systems (i.e., five sets for each aquifer system).
The results show that ACO outperforms NLP and graphical methods for all three aquifer
Ac
systems and thus is a more reliable technique for estimating the hydraulic parameters of confined,
unconfined, and leaky aquifers. On average, for the five investigated confined aquifers, the RMSE of
estimated drawdown from ACO is 11% and 33% less than the RMSEs respectively from the NLP and
graphical methods. These values are 28% and 54% for the five unconfined aquifers and 40% and 49%
for the leaky aquifers. Similarly, GA performs better than the NLP and graphical methods. For the five
confined aquifers, on average, the drawdown estimates from ACO have a RMSE of 0.0772 m, which
is a 22 % decrease of the RMSE of 0.0991 m from the graphical method and a 9 % reduction of the
19
Page 19 of 33
RMSE of 0.0844 m from NLP. The results from ACO and GA are close and comparable. Results also
suggest that the errors associated with the NLP approach are less than those of graphical method and
thus NLP can estimate the hydraulic parameters more accurately than the graphical method.
Comparing the running time of graphical, NLP, GA, and ACO indicates that ACO converges to
t
the optimal solution faster than other methods, while the graphical approach has the lowest
ip
convergence rate.
cr
With the presence of large memory and high speed computers nowadays, the use of ACO
us
approach is strongly recommended for estimating hydraulic parameters of aquifers from the pumping
test data in lieu of the subjective and burdensome graphical and NLP methods.
References
[1]
an
K.S. Balkhair, Aquifer parameters determination for large diameter wells using neural network
M
approach, Journal of Hydrology, 265 (2002) 118-128.
[2] C.R. Fitts, Groundwater Science, Elsevier, Academic Press, 2002.
[3] Y.-C. Huang, H.-D. Yeh, The use of sensitivity analysis in on-line aquifer parameter
d
20
Page 20 of 33
[9] D. Ahlfeld, G. Baro-Montes, Solving Unconfined Groundwater Flow Management Problems
with Successive Linear Programming, Journal of Water Resources Planning and Management,
134 (2008) 404-412.
[10] C.M. Colson, M.H. Nehrir, S.A. Pourmousavi, Towards real-time microgrid power
management using computational intelligence methods, in: Power and Energy Society General
t
Meeting, 2010 IEEE, 2010, pp. 1-8.
ip
[11] F.A.C. Viana, G.I. Kotinda, D.A. Rade, V. Steffen Jr, Tuning dynamic vibration absorbers by
using ant colony optimization, Computers & Structures, 86 (2008) 1539-1549.
cr
[12] K. El Harrouni, D. Ouazar, G.A. Walters, A.H.D. Cheng, Groundwater optimization and
parameter estimation by genetic algorithm and dual reciprocity boundary element method,
us
Engineering Analysis with Boundary Elements, 18 (1996) 287-296.
[13] S. Lingireddy, Aquifer Parameter Estimation Using Genetic Algorithms and Neural Networks,
an
Civil Engineering and Environmental Systems, 15 (1998) 125-144.
[14] M. Rajesh, D. Kashyap, K.S. Hari Prasad, Estimation of unconfined aquifer parameters by
genetic algorithms, Hydrological Sciences Journal, 55 (2010) 403-413.
M
[15] M. López-Ibáñez, T. Prasad, B. Paechter, Ant Colony Optimization for Optimal Control of
Pumps in Water Distribution Networks, Journal of Water Resources Planning and
Management, 134 (2008) 337-346.
d
[16] H.R. Maier, A.R. Simpson, A.C. Zecchin, W.K. Foong, K.Y. Phang, H.Y. Seah, C.L. Tan, Ant
te
Colony Optimization for Design of Water Distribution Systems, Journal of Water Resources
Planning and Management, 129 (2003) 200-209.
p
[17] A. Ostfeld, A. Tubaltzev, Ant Colony Optimization for Least-Cost Design and Operation of
ce
Pumping Water Distribution Systems, Journal of Water Resources Planning and Management,
134 (2008) 107-118.
[18] A.C. Zecchin, A.R. Simpson, H.R. Maier, J.B. Nixon, Parametric study for an ant algorithm
Ac
21
Page 21 of 33
[21] D.N. Kumar, M.J. Reddy, Ant Colony Optimization for Multi-Purpose Reservoir Operation,
Water Resources Management, 20 (2006) 879-898.
[22] R. Moeini, M.H. Afshar, Application of an Ant Colony Optimization Algorithm for Optimal
Operation of Reservoirs: A Comparative Study of Three Proposed Formulations, Transaction
A: Civil Engineering, 16 (2009) 273-285.
t
[23] V. Nourani, S. Talatahari, P. Monadjemi, S. Shahradfar, Application of ant colony
ip
optimization to optimal design of open channels, Journal of Hydraulic Research, 47 (2009)
656-665.
cr
[24] X. He, J.J. Liu, Aquifer Parameter Identification with Ant Colony Optimization Algorithm, in:
Intelligent Systems and Applications, 2009. ISA 2009. International Workshop on, 2009, pp. 1-
us
4.
[25] B. Ataie-Ashtiani, H. Ketabchi, Elitist Continuous Ant Colony Optimization Algorithm for
an
Optimal Management of Coastal Aquifers, Water Resources Management, 25 (2011) 165-190.
[26] S. Li, Y. Liu, H. Yu, Parameter Estimation Approach in Groundwater Hydrology Using Hybrid
Ant Colony System, in: D.-S. Huang, K. Li, G. Irwin (Eds.) Computational Intelligence and
M
Bioinformatics, Springer Berlin Heidelberg, 2006, pp. 182-191.
[27] V. Batu, Aquifer Hydraulics: A Comprehensive Guide to Hydrogeologic Data Analysis, Wiley,
1998.
d
[28] R.J. Charbeneau, Groundwater Hydraulics And Pollutant Transport, Waveland Press,
te
Incorporated, 2006.
[29] C.W. Fetter, Applied Hydrogeology, Charles E. Merrill Publishing Company, 1980.
p
[30] H.M. Raghunath, Ground Water, New Age International (P) Limited, 2007.
ce
[31] D.K. Todd, GROUNDWATER HYDROLOGY, 2ND ED, Wiley India Pvt. Limited, 1980.
[32] W.C. Walton, Groundwater resource evaluation, McGraw-Hill, 1970.
[33] W.C. Walton, Aquifer Test Analysis with WindowsTM Software, CRC/Lewis Publishers,
Ac
1996.
[34] C.V. Theis, The relation between the lowering of the Piezometric surface and the rate and
duration of discharge of a well using ground-water storage, Transactions American
Geophysical Union, 16 (1935) 519-524.
[35] P.-H. Tseng, T.-C. Lee, Numerical evaluation of exponential integral: Theis well function
approximation, Journal of Hydrology, 205 (1998) 38-51.
[36] C.E. Jacob, Notes on Determining Permeability by Pumping Tests Under Water-table
Conditions, The Author, 1944.
22
Page 22 of 33
[37] S.P. Neuman, Analysis of pumping test data from anisotropic unconfined aquifers considering
delayed gravity response, Water Resources Research, 11 (1975) 329-342.
[38] M.S. Hantush, C.E. Jacob, Non-steady radial flow in an infinite leaky aquifer, Transactions
American Geophysical Union, 36 (1955) 95-100.
[39] M.S. Hantush, Analysis of data from pumping tests in leaky aquifers, Transactions American
t
Geophysical Union, 37 (1956) 702-714.
ip
[40] M.S. Hantush, Modification of the theory of leaky aquifers, Journal of Geophysical Research,
65 (1960) 3713-3725.
cr
[41] A. Brooke, D. Kendrick, A. Meeraus, GAMS: release 2.25 : a user's guide, Scientific Press,
1992.
us
[42] M. Ahmad, G.P. Kutcher, Irrigation planning with environmental considerations : a case study
of Pakistan's Indus Basin / Masood Ahmad and Gary P. Kutcher, World Bank, 1992.
an
[43] N. Alam, T. Olsthoorn, Sustainable conjunctive use of surface and groundwater: Modeling on
the basin scale, International journal of natural resources and marine sciences, (2011).
[44] K.C. Knapp, M. Weinberg, R. Howitt, J.F. Posnikoff, Water transfers, agriculture, and
M
groundwater management: a dynamic economic analysis, Journal of Environmental
Management, 67 (2003) 291-301.
[45] A.M. Michelsen, R.A. Young, Optioning Agricultural Water Rights for Urban Water Supplies
d
[48] A.S. Drud, CONOPT: A System for Large-Scale Nonlinear Optimization, Reference Manual
for CONOPT Subroutine Library, in, ARKI Consulting and Development A/S, Bagsvaerd,
Denmark, 1996.
[49] D.E. Goldberg, Genetic Algorithms in Search, Optimization, and Machine Learning, Addison-
Wesley, 1989.
[50] Z. Michalewicz, Genetic Algorithms + Data Structures = Evolution Programs, Springer, 1996.
[51] M. Dorigo, V. Maniezzo, A. Colorni, Ant system: optimization by a colony of cooperating
agents, IEEE Transactions on Systems, Man, and Cybernetics–Part B, 26 (1996) 29-41.
[52] M. Dorigo, T. Stutzle, Ant Colony Optimization, The MIT Press, 2004.
23
Page 23 of 33
[53] K.C. Abbaspour, R. Schulin, M.T. van Genuchten, Estimating unsaturated soil hydraulic
parameters using ant colony optimization, Advances in Water Resources, 24 (2001) 827-841.
[54] M. Afshar, R. Moeini, Partially and Fully Constrained Ant Algorithms for the Optimal
Solution of Large Scale Reservoir Operation Problems, Water Resources Management, 22
(2008) 1835-1857.
t
[55] M. Jalali, A. Afshar, M. Mariño, Multi-Colony Ant Algorithm for Continuous Multi-Reservoir
ip
Operation Optimization Problem, Water Resources Management, 21 (2007) 1429-1447.
[56] S.P. Neuman, A. Blattstein, M. Riva, D.M. Tartakovsky, A. Guadagnini, T. Ptak, Type curve
cr
interpretation of late-time pumping test data in randomly heterogeneous aquifers, Water
Resources Research, 43 (2007) W10421.
us
[57] P. Trinchero, X. Sanchez-Vila, N. Copty, A. Findikakis, A new method for the interpretation of
pumping tests in leaky aquifers, Ground Water, 46 (2008) 133-143.
an
M
d
p te
ce
Ac
24
Page 24 of 33
Figure captions:
Figure 1. Flowchart of GA method.
t
ip
Figure 4. ACO algorithm for identifying hydraulic parameters of an aquifer.
Figure 5. Comparison of time-drawdown plots from graphical, NLP, and ACO techniques for sample
cr
confined (datasets C3 and C4), unconfined (datasets U1, U2, and U4), and leaky (dataset L1) aquifers.
us
Table captions:
an
Table 1. Comparison of estimated hydraulic parameters from graphical, NLP, GA, and ACO
techniques for confined aquifers. The best results and its associated method are bolded.
Table 2. Comparison of estimated hydraulic parameters from graphical, NLP, GA, and ACO
M
techniques for unconfined aquifers. The best results and its associated method are bolded.
Table 3. Comparison of estimated hydraulic parameters from graphical, NLP, GA, and ACO
d
techniques for leaky aquifers. The best results and its associated method are bolded.
te
Table 4. Comparing running times of graphical, NLP, GA and ACO techniques for various aquifer
types. The lowest running time and its associated approach are bolded.
p
ce
Ac
25
Page 25 of 33
t
ip
cr
us
an
M
d
te
p
ce
Ac
26
Page 26 of 33
t
ip
cr
us
an
Figure 2. Schematic illustration of one point crossover
M
d
p te
ce
Ac
27
Page 27 of 33
t
ip
cr
us
an
Figure 3. Schematic of ant routes traveling between five discrete parameters.
M
d
p te
ce
Ac
28
Page 28 of 33
t
ip
Start
cr
Initiate ACO System
us
values for unknown aquifer parameters)
an
Calculate drawdown (Sc)
M
Calculate Objective Function
g = ∑ (S c − S o )
2
Iter. No. =
d
Iter. No + 1
Output
te
No
ce
29
Page 29 of 33
t
ip
cr
us
an
M
d
p te
ce
Ac
Figure 5. Comparison of time-drawdown plots from graphical, NLP, and ACO techniques for sample confined
(datasets C3 and C4), unconfined (datasets U1, U2, and U4), and leaky (dataset L1) aquifers.
30
Page 30 of 33
Table 1. Comparison of estimated hydraulic parameters from graphical, NLP, GA, and ACO
techniques for confined aquifers. The best results and its associated method are bolded.
t
2
Coefficient (m /day) (m) (m)
ip
Graphical 0.002214 2090 0.2441 0.3184 0.9251
Confined1 NLP 0.001240 2285 0.2766 0.2931 0.9215
cr
(C1) GA 0.001998 2195 0.2216 0.2561 0.9300
ACO 0.001860 2021 0.2372 0.2737 0.9270
us
Graphical 0.000203 237 0.0586 0.0676 0.9861
Confined2 NLP 0.000174 234 0.0402 0.0457 0.9865
(C2) GA 0.000166 235 0.0395 0.0454 0.9871
ACO 0.000156 236 0.0388 0.0445 0.9868
an
Graphical 0.000307 256 0.0441 0.0490 0.9868
Confined3 NLP 0.000287 245 0.0324 0.0398 0.9873
(C3) GA 0.000289 247 0.0315 0.0378 0.9882
M
ACO 0.000287 246 0.0321 0.0398 0.9877
Graphical 0.002993 13929 0.0338 0.0521 0.8854
Confined4 NLP 0.001819 12096 0.0293 0.0382 0.9132
d
(C4) GA 0.003891 7819 0.0193 0.0257 0.9324
ACO 0.005302 5559 0.0163 0.0230 0.9464
te
31
Page 31 of 33
Table 2. Comparison of estimated hydraulic parameters from graphical, NLP, GA, and ACO
techniques for confined aquifers. The best results and its associated method are bolded.
Hydraulic
Specific MAE RMSE
Dataset Method Conductivity R2
Yield (m) (m)
(m/day)
t
ip
Graphical 0.02492 80 0.0493 0.0598 0.9665
Unconfined1 NLP 0.01627 116 0.0296 0.0323 0.9672
GA 0.02912 71 0.0255 0.0300 0.9701
cr
(U1)
ACO 0.02476 78 0.0167 0.0194 0.9792
Graphical 0.01036 55 0.1475 0.1724 0.8309
us
Unconfined2 NLP 0.00474 58 0.1496 0.1704 0.8023
(U2) GA 0.07532 31 0.0489 0.0671 0.9724
ACO 0.06928 23 0.0646 0.0801 0.9616
an
Graphical 0.00687 200 0.0402 0.0506 0.9463
Unconfined3 NLP 0.00491 183 0.0288 0.0322 0.9525
(U3) GA 0.00551 195 0.2880 0.0320 0.9571
M
ACO 0.00651 169 0.0283 0.0314 0.9582
Graphical 0.06280 156 0.0507 0.0755 0.9713
Unconfined4 NLP 0.03771 150 0.0390 0.0472 0.9725
(U4) GA 0.04619 134 0.0317 0.0403 0.9906
d
32
Page 32 of 33
Table 3. Comparison of estimated hydraulic parameters from graphical, NLP, GA and ACO
techniques for leaky aquifers. The best results and its associated method are bolded.
t
Leaky1 NLP 0.0000674 704.8 933.4 0.0058 0.0066 0.9963
ip
(L1) GA 0.0000901 820.9 1341.6 0.0027 0.0035 0.9991
ACO 0.0000814 772.4 1692.0 0.0033 0.0042 0.9987
cr
Graphical 0.0000140 1045.4 9985.2 0.0054 0.0072 0.9979
Leaky2 NLP 0.0000477 793.7 2500.0 0.0056 0.0065 0.9968
(L2) GA 0.0000040 1001.5 7560.3 0.0051 0.0069 0.9974
us
ACO 0.0000225 960.4 9002.6 0.0039 0.0047 0.9983
Graphical 0.0010300 181.4 159.9 0.0619 0.0866 0.9881
Leaky3 NLP 0.0009842 228.7 408.7 0.0644 0.0878 0.9852
an
(L3) GA 0.0011000 210.3 345.1 0.0588 0.0765 0.9804
ACO 0.0010460 183.1 221.0 0.0401 0.0519 0.9940
Graphical 0.0001960 22.9 192.9 0.0557 0.0719 0.9946
Leaky4
M
NLP 0.0001196 29.1 474.5 0.0358 0.0422 0.9956
(L4) GA 0.0001256 27.6 341.1 0.0267 0.0356 0.9971
ACO 0.0001377 25.2 250.8 0.0197 0.0255 0.9984
Graphical 0.0001480 26.4 292.1 0.0515 0.06985 0.9931
d
Table 4. Comparing running times of graphical, NLP, GA and ACO techniques for various aquifer
ce
types. The lowest running time and its associated approach are bolded.
Graphical 18.9
NLP 11.1
Confined (C1-C5)
GA 13.0
ACO 8.2
Graphical 20.0
NLP 11.8
Unconfined (U1-U5)
GA 14.0
ACO 9.4
Graphical 46.4
NLP 25.3
Leaky (L1-L5)
GA 28.3
ACO 16.8
33
Page 33 of 33