Abstract
In this study, we present a comparative assessment of simulation-optimization (S-O) models to estimate aquifer parameters such as transmissivity, longitudinal dispersivity, and transverse dispersivity. The groundwater flow and contaminant transport processes are simulated using the mesh-free radial basis point collocation method (RPCM). Four different S-O models are developed by combining the RPCM model separately with genetic algorithm (GA), differential evolution (DE), cat swarm optimization (CSO), and particle swarm optimization (PSO). The objective of the S-O model is to minimize a composite objective function with transmissivity, longitudinal dispersivity, and transverse dispersivity as decision variables. Hydraulic head and contaminant concentration at observation points are the state variables. The S-O models are used to estimate aquifer parameters of a confined aquifer with nine zones. It is found that RPCM-based DE, CSO, and PSO models are more accurate in estimating aquifer parameters than RPCM-GA. However, for noisy observed data, the RPCM-CSO model outperforms other models. The efficiency of the RPCM-CSO model over other models is further established by performing reliability analysis to the noisy observed data set. The comparative study reflects the efficacy of CSO over GA, DE, and PSO.
INTRODUCTION
Estimation of aquifer parameters such as storativity, transmissivity, porosity, and longitudinal dispersivity by the inverse modeling approach is very important for groundwater management (Peralta 2012). In groundwater inverse modeling, transmissivity, storativity, recharge, leakage, longitudinal dispersivity, transverse dispersivity, and molecular diffusion are the decision variables, whereas the hydraulic head and contaminant concentrations at predefined locations are the state variables (Medina & Carrera 1996). The quality of groundwater models cannot be simply improved by increasing the accuracy of forward solutions if inverse solutions are not properly solved (Sun 1994). In many previous studies, inverse modeling has been widely used to estimate aquifer parameters (Carrera & Neuman 1986a, 1986b; Zimmerman et al. 1998; Prasad & Rastogi 2001; Rajanayaka & Kulasiri 2001; Huggi & Rastogi 2003; Carrera et al. 2005; Franssen et al. 2009; Jin et al. 2009; Wang et al. 2017). The quantity and quality of the observed spatiotemporal groundwater data (hydraulic head and contaminant concentrations at observation location) greatly influence the results of inverse modeling and most of the time the data are generally inadequate for accurate prediction (Medina & Carrera 1996). Another major thing that governs the error in inverse modeling is inaccurate model structure identification. By only changing performance criteria and optimization techniques, satisfactory results cannot be found for inverse modeling (Sun 2013). The inverse model determines aquifer parameters by using the spatiotemporal observed data.
Numerical modeling tools such as the finite difference method (FDM), finite element method (FEM), and analytical element method (AEM) are mostly used for groundwater flow and contaminant transport simulation (Fitts 2002). The FDM and FEM require preprocessing effort for mesh generation, hence they are computationally expensive for large aquifers. The AEM has limitations in simulating a highly heterogeneous medium and transient flow condition (Majumder & Eldho 2017). On the contrary, many studies claimed that mesh-free methods are computationally more efficient and accurate than FDM and FEM in simulating large-scale groundwater flow and contaminant transport problems (Guneshwor et al. 2018; Praveen Kumar & Dodagoudar 2008; Kovářík & Mužík 2013; Meenal & Eldho 2012; Patel & Rastogi 2017; Swathi & Eldho 2018; Thomas et al. 2014, 2018).
Heuristic optimizers such as GA, simulated annealing (SA), DE, particle swarm optimization (PSO), and ant colony optimization (ACO) were mostly used to estimate aquifer parameters by inverse modeling by earlier researchers (Abbaspour et al. 2001; Prasad & Rastogi 2001; Reed et al. 2001; Huggi & Rastogi 2003; Li et al. 2006; Fonna et al. 2013; Elçi & Ayvaz 2014; Gurarslan & Karahan 2015; Mategaonkar et al. 2018; Swathi & Eldho 2018; Thomas et al. 2018). Advantages of using heuristic optimizers over the traditional gradient-based optimization algorithms are that they do not require derivative computations and initial point to initiate the search operation (Majumder & Eldho 2016). A comparison carried out by Matott et al. (2006) among various optimization techniques for groundwater remediation studies shows the superiority of PSO over GA, random search algorithm (RSA), SSA, and conjugate gradient. However, PSO sometimes converges prematurely. In PSO, all swarms move towards the best solution (which may be the local best solution) which is found by a specific swarm and if the velocities of all the swarms are very low, then swarms cannot come out from the point. Thus, in such cases, premature convergence may take place and the associated error in the solution is known as stagnation point error (Saha et al. 2013; Guo et al. 2016; Majumder & Eldho 2016). Recently, cat swarm optimization (CSO) has been found to perform better when compared to PSO, GA, and SA for groundwater management problems (Majumder & Eldho 2016). Seeking mode and tracking mode are the search operations that take place in CSO. In seeking mode, each cat creates many copies around its surroundings, and there is randomness associated with each copy. Hence, the search space of CSO is vast. Due to this seeking mode, the premature convergence/stagnation point error is highly unlikely in CSO (Majumder & Eldho 2016). Successful application of CSO in estimating transmissivity values in a heterogeneous confined aquifer can be found in Thomas et al. (2018).
In this study, four S-O models are developed by coupling RPCM separately with GA, DE, CSO, and PSO. The models are applied to estimate transmissivity, longitudinal dispersivity, and transverse dispersivity of a hypothetical confined aquifer case study with nine zones. Further, a reliability analysis is performed to comprehend and intercompare the overall consistency of the output from the various models.
METHODOLOGY
and 
are the transmissivities along x and y direction, 
is the vertically averaged source or sink term, q is inflow rate, and S is storativity.
and 
are the seepage velocities along x and y direction, 
and 
are the hydraulic conductivities, and 
is the porosity of the porous medium.
is the concentration of the solute species injected at the sources which are governed by Dirac delta function, and 
is the volumetric pumping rate.
) can be expressed as (Bear & Cheng 2010):
is the Kronecker delta; 
is the transverse dispersivity (L); 
is the longitudinal dispersivity (L); 
and 
is the seepage velocity in the ith and jth directions (LT−1); 
is the seepage velocity vector; 
is the effective molecular diffusion coefficient.Mesh free radial point collocation method
is expressed as a linear combination of n radial basis function (RBF) 
(Liu & Gu 2005) as follows:
is the radial basis function (RBF); q and αc, are the shape parameters; 
, dc is the average nodal spacing for all the nodes in local support domain; (x, y) are the coordinates of the point of interest and (xi, yi) are the coordinates of any node in the support domain of the point of interest. 
is known as shape function. By using the shape function that is derived as explained in Thomas et al. (2018), the hydraulic head can be approximated as:
is global matrix (GM) for shape function; 
is GM for the seccond derivative of shape function with respect to x, and 
is GM for the second derivative of shape function with respect to y.
is the constant head boundary condition, it can be easily discretized as 
. For implementing the derivative boundary condition, a direct collocation approach has been implemented. A direct collocation method for derivative boundary condition implementation is illustrated here by using Equation (10):
is GM for shape function; 
is GM for the first derivative of shape function with respect to x; 
is GM for first derivative of shape function with respect to y; 
is GM for second derivative of shape function with respect to x; 
is GM for second derivative of shape function with respect to y.Genetic algorithm (GA)
The genetic algorithm is a metaheuristic optimization technique. Based on Darwin's survival of the fittest principle, GA finds the best and fittest solution. GA can easily find the global optimum solution of large complex nonlinear models unlike traditional optimization methods. The basic elements of GA are reproduction, crossover, and mutation. More information about GA can be found in Deb (2012). In this study, we have directly used the GA optimization toolbox which is available in MATLAB.
Differential evolution (DE)
Differential evolution is a meta-heuristic evolutionary optimization algorithm proposed by Storn & Price (1995). The basic elements of DE are mutation, crossover, and selection. The steps of DE are explained below (Storn & Price 1997).
and 
are the lower bound and upper bound of the variable 
; 
is a random number 
(0,1).
).
and F is a user-defined mutation factor 
(0, 2).
) from the target vector (
) and the mutant vector (
),
is a random number 
(0,1) and 
is the user-defined crossover probability 
(0,1);
is an integer random number 
(1,2, 3, …..D).
) and trial vector (
). For the next generation select the vector which results in the lowest fitness function value.
Step 5: Continue step 2 to step 4 until the termination criterion is satisfied.
Among the various variants of DE, DE/rand/1/bin is used in the present study.
Cat swarm optimization (CSO)
Chu & Tsai (2007) proposed a metaheuristic optimization algorithm, namely, cat swarm optimization, by imitating the food searching behavior of the cat. In CSO, food searching takes place by seeking mode and tracking mode. Seeking mode is a local search process in which the cat looks for prey with very small changes in position. There are four parameters in seeking mode. These are seeking memory pool (SMP), seeking range of selected dimensions (SRD), count of dimension to change (CDC), and self-position consideration (SPC). Tracking mode is a global search process in which the cat chases the prey with a certain velocity. The velocity decreases with respect to generation number. It also signifies that the cat is moving closer to catch the prey with respect to generation number. The velocity is also multiplied by an inertia weight (w), which varies linearly from 0.9 to 0.4 with respect to generation number (Majumder & Eldho 2016). The seeking mode and tracking mode are connected by a ratio called mixture ratio (MR). The cat mostly spends time in seeking prey and little time in chasing prey. Thus, the MR is assigned in such a way that a higher number of cats belongs to seeking mode and a lower number of cats belongs in tracking mode.
A typical CSO randomly generates the initial positions and velocities of N number of cats in the D dimensional space within the specified position and velocity bound. Thereafter, evaluation of the fitness function value for each cat is done. The cat (
) with the best fitness value (
) is found and is stored in the memory. Then, all the cats are randomly distributed in the seeking and tracking modes depending upon the MR. The results of seeking and tracking mode processes are combined, to yield new position and fitness values of all the cats in the vector form. The cat with the best fitness value (
) is found and it replaces 
, if 
is better than 
. Also 
is replaced by the respective cat. Until the termination criterion is satisfied the steps from randomly distributing cats depending upon mixture ratio to seeking mode and tracking mode are repeated. The termination criterion is generally the specified maximum number of iterations/generations (Majumder & Eldho 2016). The more detailed description and equations of CSO can be found in Thomas et al. (2018).
Particle swarm optimization (PSO)
In particle swarm optimization, particles move across the search space according to global best position of all particles and local best position of the individual particle (Eberhart & Kennedy 1995). In PSO, there is a velocity update equation and a position update equation. A typical PSO has the following steps (Robinson & Rahmat 2004):
Generate N number of particles in D dimensional space.
Randomly initialize position (
) and velocity (
) of each particle within the allowable limits.Compute fitness function values of all the particles. Find global best position (
), global best fitness function value (
), individual best position (
), individual best fitness function value (
) before starting iteration and store them. Here, t is the iteration number. Before starting iteration 
. For the first iteration 
.- Update velocity and position of each particle for the next generation using the following equations (Robinson & Rahmat 2004):(17)
Check whether velocity and position are within the allowable limits. If velocity and position are out of bounds, then set them to their allowable limits. In the above expressions, w is the inertia weight, which with respect to iteration number decreases linearly from 0.9 to 0.4;(18)
and 
is the random number between 0 and 1, 
(Robinson & Rahmat 2004). With the updated position, compute the fitness value of all the particles.
If
is better than Fgbest[t], then Fgbest[t] = Fgbest[t + 1]; also replace 
. If 
is not better than 
, then keep 
and 
unchanged.Also, for each particle, if
is better than 
, then set 
and 
; otherwise, keep 
and 
unchanged.Repeat step (iv) to step (vii), until the termination criterion is achieved. Although there are many variants of PSO, we have used standard original PSO in this study (Robinson & Rahmat 2004).
In all the optimization algorithms mentioned above, the termination criterion is a user-defined number of generations.
Verification of optimization models
local minima and it is a multimodal test function. The Michalewicz function can be expressed as (Molga & Smutnicki 2005):
The search space of the function is 
. The Michalewicz function is optimized by GA, DE, PSO, and CSO by assuming dimension
. The parameters' settings for various optimization methods are given in Table 1. The total number of generations 
is kept as 1,000. The true global minimum of the Michalewicz function for the dimension 
is −9.66015 (Molga & Smutnicki 2005). The optimum values obtained by DE and CSO are very close to the true optimum (Table 2). However, CSO takes about 600 generations to reach the global optimum value whereas DE takes almost 950 generations to reach the global optimum value (Figure 1). There are no improvements in the results of PSO and GA after 300 generations. This is because the solutions of both PSO and GA are trapped in local minima. From Table 2, it can be seen that even GA and PSO have failed to reach very close to the true optimum value even after 1,000 generations. To comprehend the computational performance of CSO, DE, PSO, and GA for the Michalewicz function, a comparative study is carried out. The computational times of the CSO, PSO, DE, and GA models are given in Table 2. As can be seen from the table, the DE is computationally most efficient followed by PSO, CSO, and GA (Table 2). In this analysis, we used a computer with a Core i7 processor, a speed of 3.1 GHz and 8 GB RAM.
Parameter settings of GA, DE, PSO, and CSO
| Optimization model | Parameter | Range of value |
|---|---|---|
| GA | Population size (N) | 40 |
| Mutation ratio (MR) | 0.01 | |
| Crossover (CR) | 0.8 | |
| DE | Population size (N) | 40 |
| Mutation factor (F) | 0.8 | |
| Crossover (CR) | 0.9 | |
| PSO | Population size (N) | 40 |
| C1 | 1.49 | |
| C2 | 1.49 | |
| w | 0.9–0.4; linearly decreasing with respect to the iteration number | |
| CSO | Population size (N) | 40 |
| MR | 0.8 | |
| SMP | 15 | |
| SRD | 0.15 | |
| CDC | 80% | |
| w | 0.9–0.4, linearly decreasing with respect to the iteration number |
| Optimization model | Parameter | Range of value |
|---|---|---|
| GA | Population size (N) | 40 |
| Mutation ratio (MR) | 0.01 | |
| Crossover (CR) | 0.8 | |
| DE | Population size (N) | 40 |
| Mutation factor (F) | 0.8 | |
| Crossover (CR) | 0.9 | |
| PSO | Population size (N) | 40 |
| C1 | 1.49 | |
| C2 | 1.49 | |
| w | 0.9–0.4; linearly decreasing with respect to the iteration number | |
| CSO | Population size (N) | 40 |
| MR | 0.8 | |
| SMP | 15 | |
| SRD | 0.15 | |
| CDC | 80% | |
| w | 0.9–0.4, linearly decreasing with respect to the iteration number |
Minimum values of Michalewicz's function obtained by GA, DE, CSO, and PSO
| GA | DE | PSO | CSO | True value | |
|---|---|---|---|---|---|
| Fitness value of Michalewicz function | −8.02 | −9.621 | −8.11 | −9.632 | − 9.66015 |
| Computational time (seconds) | 21.52 | 5.12 | 6.31 | 17.78 |
| GA | DE | PSO | CSO | True value | |
|---|---|---|---|---|---|
| Fitness value of Michalewicz function | −8.02 | −9.621 | −8.11 | −9.632 | − 9.66015 |
| Computational time (seconds) | 21.52 | 5.12 | 6.31 | 17.78 |
Best fitness value versus number of generations using GA, DE, CSO, and PSO for Michalewicz's function.
Best fitness value versus number of generations using GA, DE, CSO, and PSO for Michalewicz's function.
GROUNDWATER INVERSE MODEL DEVELOPMENT
is observed hydraulic head; 
is the simulated hydraulic head; 
is observed contaminant concentration; 
is simulated contaminant concentration; L is the total number of monitoring wells; to and tf are beginning and ending time of observations. Ti is transmissivity at node i; w and W are the weighting coefficients. The first term of the composite objective function is the summation of weighted squared differences of observed and simulated hydraulic head, and the second term is the weighted squared differences of observed and simulated contaminant concentrations.Initially, the MFree-based RPCM model is developed to simulate groundwater flow and contaminant transport processes. By coupling the RPCM model separately with GE, DA, CSO, and PSO, four inverse models are developed for aquifer parameter estimation. The flowchart for the RPCM-DE model is shown in Figure 2. The flowchart of RPCM-CSO and RPCM-PSO models can be found in Thomas et al. (2018). The coupled RPCM-GA, RPCM-DE, RPCM-CSO, and RPCM-PSO models are applied to estimate the aquifer parameters of a hypothetical confined aquifer by minimizing Equation (20).
Simulation optimization flowchart using RPCM-DE.
MODEL APPLICATION AND ANALYSIS
The RPCM-GA, RPCM-CSO, RPCM-DE, and RPCM-PSO models are applied to estimate the aquifer parameters of a hypothetical confined aquifer with an area of 36 sq. km (Rastogi & Huggi 2009). The aquifer domain is shown in Figure 3. The southern boundary is a constant head boundary of 100 m. The western boundary is a constant flux boundary with an inflow rate of 0.25 m2/day. Both the northern and eastern boundaries are no-flow boundaries. Two aquitard recharge zones exist at the north with recharge at a rate of 0.00015 m/day and 0.00025 m/day, respectively. Also total dissolved solids (TDS) concentration in the two recharge zones are 80 ppm and 100 ppm, respectively. Table 3 gives the true values of transmissivity, longitudinal dispersivity, and transverse dispersivity values of the nine-zoned aquifer. The study area has a recharge well with a flow rate of 500 m3/d. TDS concentration of injected water in the recharge well is 1,000 ppm. There is also one pumping well with a flow rate of 1,200 m3/d. As shown in Figure 4, 18 observations nodes (wells) are chosen within the flow region. The storage coefficient of the aquifer is assumed as 0.001. Transmissivities vary from 5 to 150 m2/day for the nine zones. The flow and transport processes of the aquifer are simulated using RPCM by scattering 49 field nodes in the aquifer domain (Figure 4).
True aquifer parameters for various zones
| Zone | Transmissivity value (m2/d) | Longitudinal dispersivity (m) | Transverse dispersivity (m) |
|---|---|---|---|
| 1 | 150 | 60 | 6 |
| 2 | 150 | 60 | 6 |
| 3 | 50 | 40 | 4 |
| 4 | 150 | 60 | 6 |
| 5 | 50 | 40 | 4 |
| 6 | 15 | 15 | 1.5 |
| 7 | 50 | 40 | 4 |
| 8 | 15 | 15 | 1.5 |
| 9 | 5 | 10 | 1 |
| Zone | Transmissivity value (m2/d) | Longitudinal dispersivity (m) | Transverse dispersivity (m) |
|---|---|---|---|
| 1 | 150 | 60 | 6 |
| 2 | 150 | 60 | 6 |
| 3 | 50 | 40 | 4 |
| 4 | 150 | 60 | 6 |
| 5 | 50 | 40 | 4 |
| 6 | 15 | 15 | 1.5 |
| 7 | 50 | 40 | 4 |
| 8 | 15 | 15 | 1.5 |
| 9 | 5 | 10 | 1 |
Aquifer domain and boundary conditions.
Nodal distribution and zonation pattern of the aquifer domain.
A sensitivity analysis with three different q values mentioned in Liu & Gu (2005) is conducted to know the best parameter suited for the problem attempted with values of 0.5, 0.98, and 1.03. It has been observed that for the particular problem, 
= 0.98 gives better results (Tables 4 and 5). It is also observed that all three values do not significantly affect the accuracy of the solution. Mean absolute percentage error (MAPE) of hydraulic head and contaminant concentration estimated with different q values are given in Table 6. Since MAPE of both estimated hydraulic head and concentration was lower with the shape parameter q = 0.98, the present study is carried out with q as 0.98. The other shape parameter of the RBF is 
. The solution accuracy is found to depend on 
as it has a profound impact on the quality of the interpolation within a support domain. Although the choice of the optimum value of this parameter is crucial, there is no established method for finding the optimum value of this parameter. In this study, it is observed that for 
, the result obtained by RPCM is in close agreement with FEM (Rastogi & Huggi 2009).
Hydraulic head (in meter) obtained by FEM and RPCM for different q values
| Node number | Head (FEM)) | Head RPCM (q = 0.5) | Absolute % difference | Head RPCM (q = 0.98) | Absolute % difference | Head RPCM (q = 1.03) | Absolute % difference |
|---|---|---|---|---|---|---|---|
| 2 | 103.16 | 101.73 | 1.386 | 102.17 | 0.960 | 102.13 | 0.998 |
| 23 | 100.30 | 99.84 | 0.459 | 100.56 | 0.259 | 100.27 | 0.030 |
| 37 | 99.30 | 98.69 | 0.614 | 99.89 | 0.594 | 99.76 | 0.463 |
| 17 | 102.47 | 101.31 | 1.132 | 103.47 | 0.976 | 103.29 | 0.800 |
| 31 | 98.92 | 94.35 | 4.620 | 95.40 | 3.558 | 95.27 | 3.690 |
| 45 | 98.15 | 99.17 | 1.039 | 99.26 | 1.131 | 99.68 | 1.559 |
| 11 | 105.35 | 100.31 | 4.784 | 100.01 | 5.069 | 100.75 | 4.366 |
| 25 | 98.15 | 97.16 | 1.009 | 98.75 | 0.611 | 98.86 | 0.723 |
| 39 | 93.62 | 99.21 | 5.971 | 98.03 | 4.711 | 97.91 | 4.582 |
| 5 | 108.34 | 106.85 | 1.375 | 107.22 | 1.034 | 107.08 | 1.163 |
| 19 | 104.04 | 108.67 | 4.450 | 107.56 | 3.383 | 108.28 | 4.075 |
| 47 | 91.26 | 91.42 | 0.175 | 91.65 | 0.427 | 92.49 | 1.348 |
| 6 | 111.65 | 110.21 | 1.290 | 110.84 | 0.725 | 111.75 | 0.090 |
| 20 | 105.16 | 107.55 | 2.273 | 107.92 | 2.625 | 108.12 | 2.815 |
| 34 | 91.18 | 93.47 | 2.512 | 93.26 | 2.281 | 93.89 | 2.972 |
| 14 | 108.49 | 103.17 | 4.904 | 103.42 | 4.673 | 103.71 | 4.406 |
| 28 | 98.69 | 97.71 | 0.993 | 98.50 | 0.193 | 99.16 | 0.476 |
| 49 | 91.74 | 91.58 | 0.174 | 91.83 | 0.098 | 92.04 | 0.327 |
| Node number | Head (FEM)) | Head RPCM (q = 0.5) | Absolute % difference | Head RPCM (q = 0.98) | Absolute % difference | Head RPCM (q = 1.03) | Absolute % difference |
|---|---|---|---|---|---|---|---|
| 2 | 103.16 | 101.73 | 1.386 | 102.17 | 0.960 | 102.13 | 0.998 |
| 23 | 100.30 | 99.84 | 0.459 | 100.56 | 0.259 | 100.27 | 0.030 |
| 37 | 99.30 | 98.69 | 0.614 | 99.89 | 0.594 | 99.76 | 0.463 |
| 17 | 102.47 | 101.31 | 1.132 | 103.47 | 0.976 | 103.29 | 0.800 |
| 31 | 98.92 | 94.35 | 4.620 | 95.40 | 3.558 | 95.27 | 3.690 |
| 45 | 98.15 | 99.17 | 1.039 | 99.26 | 1.131 | 99.68 | 1.559 |
| 11 | 105.35 | 100.31 | 4.784 | 100.01 | 5.069 | 100.75 | 4.366 |
| 25 | 98.15 | 97.16 | 1.009 | 98.75 | 0.611 | 98.86 | 0.723 |
| 39 | 93.62 | 99.21 | 5.971 | 98.03 | 4.711 | 97.91 | 4.582 |
| 5 | 108.34 | 106.85 | 1.375 | 107.22 | 1.034 | 107.08 | 1.163 |
| 19 | 104.04 | 108.67 | 4.450 | 107.56 | 3.383 | 108.28 | 4.075 |
| 47 | 91.26 | 91.42 | 0.175 | 91.65 | 0.427 | 92.49 | 1.348 |
| 6 | 111.65 | 110.21 | 1.290 | 110.84 | 0.725 | 111.75 | 0.090 |
| 20 | 105.16 | 107.55 | 2.273 | 107.92 | 2.625 | 108.12 | 2.815 |
| 34 | 91.18 | 93.47 | 2.512 | 93.26 | 2.281 | 93.89 | 2.972 |
| 14 | 108.49 | 103.17 | 4.904 | 103.42 | 4.673 | 103.71 | 4.406 |
| 28 | 98.69 | 97.71 | 0.993 | 98.50 | 0.193 | 99.16 | 0.476 |
| 49 | 91.74 | 91.58 | 0.174 | 91.83 | 0.098 | 92.04 | 0.327 |
Contaminant concentration (in ppm) obtained by FEM and RPCM for different q values
| Node number | Concentration (FEM) | Concentration RPCM (q = 0.5) | Absolute % difference | Concentration RPCM (q = 0.98) | Absolute % difference | Concentration RPCM (q = 1.03) | Absolute % difference |
|---|---|---|---|---|---|---|---|
| 2 | 101.55 | 105.94 | 4.323 | 105.42 | 3.811 | 105.73 | 4.116 |
| 23 | 171.99 | 177.67 | 3.303 | 177.44 | 3.169 | 177.49 | 3.198 |
| 37 | 101.11 | 104.64 | 3.491 | 104.42 | 3.274 | 104.48 | 3.333 |
| 17 | 456.00 | 448.97 | 1.542 | 448.94 | 1.548 | 448.92 | 1.553 |
| 31 | 197.00 | 193.91 | 1.569 | 193.74 | 1.655 | 193.83 | 1.609 |
| 45 | 100.07 | 104.97 | 4.897 | 104.56 | 4.487 | 104.84 | 4.767 |
| 11 | 129.76 | 132.84 | 2.374 | 132.47 | 2.088 | 132.61 | 2.196 |
| 25 | 833.44 | 838.24 | 0.576 | 837.65 | 0.505 | 837.98 | 0.545 |
| 39 | 88.13 | 90.97 | 3.223 | 90.43 | 2.610 | 90.53 | 2.723 |
| 5 | 114.58 | 118.15 | 3.116 | 117.47 | 2.522 | 117.88 | 2.880 |
| 19 | 343.92 | 351.90 | 2.320 | 351.54 | 2.216 | 351.54 | 2.216 |
| 47 | 101.39 | 102.94 | 1.529 | 102.47 | 1.065 | 102.62 | 1.213 |
| 6 | 113.55 | 117.76 | 3.708 | 117.24 | 3.250 | 117.53 | 3.505 |
| 20 | 66.43 | 68.94 | 3.778 | 68.67 | 3.372 | 68.81 | 3.583 |
| 34 | 75.17 | 79.27 | 5.454 | 78.16 | 3.978 | 78.11 | 3.911 |
| 14 | 104.26 | 108.42 | 3.990 | 107.55 | 3.156 | 108.16 | 3.741 |
| 28 | 102.06 | 106.34 | 4.194 | 105.49 | 3.361 | 106.04 | 3.900 |
| 49 | 100.00 | 100.00 | 0.000 | 100.00 | 0.000 | 100.00 | 0.000 |
| Node number | Concentration (FEM) | Concentration RPCM (q = 0.5) | Absolute % difference | Concentration RPCM (q = 0.98) | Absolute % difference | Concentration RPCM (q = 1.03) | Absolute % difference |
|---|---|---|---|---|---|---|---|
| 2 | 101.55 | 105.94 | 4.323 | 105.42 | 3.811 | 105.73 | 4.116 |
| 23 | 171.99 | 177.67 | 3.303 | 177.44 | 3.169 | 177.49 | 3.198 |
| 37 | 101.11 | 104.64 | 3.491 | 104.42 | 3.274 | 104.48 | 3.333 |
| 17 | 456.00 | 448.97 | 1.542 | 448.94 | 1.548 | 448.92 | 1.553 |
| 31 | 197.00 | 193.91 | 1.569 | 193.74 | 1.655 | 193.83 | 1.609 |
| 45 | 100.07 | 104.97 | 4.897 | 104.56 | 4.487 | 104.84 | 4.767 |
| 11 | 129.76 | 132.84 | 2.374 | 132.47 | 2.088 | 132.61 | 2.196 |
| 25 | 833.44 | 838.24 | 0.576 | 837.65 | 0.505 | 837.98 | 0.545 |
| 39 | 88.13 | 90.97 | 3.223 | 90.43 | 2.610 | 90.53 | 2.723 |
| 5 | 114.58 | 118.15 | 3.116 | 117.47 | 2.522 | 117.88 | 2.880 |
| 19 | 343.92 | 351.90 | 2.320 | 351.54 | 2.216 | 351.54 | 2.216 |
| 47 | 101.39 | 102.94 | 1.529 | 102.47 | 1.065 | 102.62 | 1.213 |
| 6 | 113.55 | 117.76 | 3.708 | 117.24 | 3.250 | 117.53 | 3.505 |
| 20 | 66.43 | 68.94 | 3.778 | 68.67 | 3.372 | 68.81 | 3.583 |
| 34 | 75.17 | 79.27 | 5.454 | 78.16 | 3.978 | 78.11 | 3.911 |
| 14 | 104.26 | 108.42 | 3.990 | 107.55 | 3.156 | 108.16 | 3.741 |
| 28 | 102.06 | 106.34 | 4.194 | 105.49 | 3.361 | 106.04 | 3.900 |
| 49 | 100.00 | 100.00 | 0.000 | 100.00 | 0.000 | 100.00 | 0.000 |
Mean absolute percentage error of head and concentration estimation with different q values using RPCM
| Shape parameter (q) value | Mean absolute % error for head | Mean absolute % error for concentration |
|---|---|---|
| 0.5 | 2.176 | 2.966 |
| 0.98 | 1.850 | 2.559 |
| 1.03 | 1.938 | 2.722 |
| Shape parameter (q) value | Mean absolute % error for head | Mean absolute % error for concentration |
|---|---|---|
| 0.5 | 2.176 | 2.966 |
| 0.98 | 1.850 | 2.559 |
| 1.03 | 1.938 | 2.722 |
The time step considered for the flow and transport simulation is considered to be 1 day. The groundwater head and concentration values obtained using the RPCM model for the known transmissivity and dispersivity values for the flow and transport problem are taken as observed head and concentration in inverse modeling. The flow and transport simulations are carried out by RPCM for a total duration of 1,000 days. The head and concentration plots after 1,000 days are plotted in Figures 5 and 6. Comparison of hydraulic head and concentration distribution obtained after 1,000 days using RPCM and FEM are shown in Table 5. As can be seen from Table 5, the hydraulic head and contaminant concentration obtained by both the methods are in good agreement.
The contour of hydraulic head (m) after 1,000 days.
The contour of contaminant concentration (ppm) after 1,000 days.
Further, the inverse models, namely, RPCM-GA, RPCM-DE, RPCM-CSO, and RPCM-PSO are applied to the hypothetical confined aquifer. All four models can predict the aquifer parameters quite accurately (Tables 7–9). However, if we look closely, it can be found that DE, CSO, and PSO are more efficient in predicting aquifer parameters than GA.
Comparison of estimated transmissivity (
/day) values in various zones using DE, GA, PSO, and CSO
| Data set | Zone | True values (m2/d) | Estimated values using CSO (m2/d) | Absolute % error CSO | Estimated values using PSO (m2/d) | Absolute % error PSO | Estimated values using DE (m2/d) | Absolute % error DE | Estimated values using GA (m2/d) | Absolute % error GA |
|---|---|---|---|---|---|---|---|---|---|---|
| Data set 1 | 1 | 150 | 149.99 | 0.01 | 149.99 | 0.01 | 149.97 | 0.02 | 149.23 | 0.51 |
| 2 | 150 | 149.98 | 0.01 | 149.98 | 0.01 | 149.98 | 0.01 | 149.37 | 0.42 | |
| 3 | 50 | 49.98 | 0.03 | 49.90 | 0.20 | 49.96 | 0.08 | 49.34 | 1.32 | |
| 4 | 150 | 149.97 | 0.02 | 149.89 | 0.07 | 149.95 | 0.03 | 149.85 | 0.10 | |
| 5 | 50 | 49.97 | 0.06 | 50.10 | 0.20 | 49.96 | 0.08 | 49.95 | 0.10 | |
| 6 | 15 | 14.99 | 0.09 | 14.97 | 0.19 | 14.96 | 0.27 | 14.92 | 0.53 | |
| 7 | 50 | 49.98 | 0.04 | 49.98 | 0.04 | 49.97 | 0.06 | 49.48 | 1.04 | |
| 8 | 15 | 14.98 | 0.16 | 14.97 | 0.18 | 14.97 | 0.20 | 14.64 | 2.40 | |
| 9 | 5 | 4.99 | 0.20 | 4.99 | 0.28 | 4.94 | 1.20 | 4.73 | 5.40 | |
| Data set 2: μ = 0, σ = 1 | 1 | 150 | 148.93 | 0.71 | 148.79 | 0.81 | 149.14 | 0.57 | 148.712 | 0.86 |
| 2 | 150 | 149.48 | 0.35 | 148.94 | 0.71 | 148.92 | 0.72 | 149.03 | 0.65 | |
| 3 | 50 | 49.16 | 1.68 | 49.22 | 1.56 | 49.24 | 1.52 | 49.31 | 1.38 | |
| 4 | 150 | 149.07 | 0.62 | 148.91 | 0.73 | 148.74 | 0.84 | 148.36 | 1.09 | |
| 5 | 50 | 49.13 | 1.74 | 49.09 | 1.82 | 49.26 | 1.48 | 49.18 | 1.64 | |
| 6 | 15 | 14.43 | 3.80 | 14.20 | 5.33 | 14.61 | 2.60 | 13.975 | 6.83 | |
| 7 | 50 | 48.98 | 2.04 | 49.18 | 1.64 | 48.52 | 2.96 | 49.27 | 1.46 | |
| 8 | 15 | 14.66 | 2.27 | 14.37 | 4.20 | 14.67 | 2.20 | 14.16 | 5.60 | |
| 9 | 5 | 4.72 | 5.60 | 4.49 | 10.20 | 4.48 | 10.40 | 4.37 | 12.60 |
| Data set | Zone | True values (m2/d) | Estimated values using CSO (m2/d) | Absolute % error CSO | Estimated values using PSO (m2/d) | Absolute % error PSO | Estimated values using DE (m2/d) | Absolute % error DE | Estimated values using GA (m2/d) | Absolute % error GA |
|---|---|---|---|---|---|---|---|---|---|---|
| Data set 1 | 1 | 150 | 149.99 | 0.01 | 149.99 | 0.01 | 149.97 | 0.02 | 149.23 | 0.51 |
| 2 | 150 | 149.98 | 0.01 | 149.98 | 0.01 | 149.98 | 0.01 | 149.37 | 0.42 | |
| 3 | 50 | 49.98 | 0.03 | 49.90 | 0.20 | 49.96 | 0.08 | 49.34 | 1.32 | |
| 4 | 150 | 149.97 | 0.02 | 149.89 | 0.07 | 149.95 | 0.03 | 149.85 | 0.10 | |
| 5 | 50 | 49.97 | 0.06 | 50.10 | 0.20 | 49.96 | 0.08 | 49.95 | 0.10 | |
| 6 | 15 | 14.99 | 0.09 | 14.97 | 0.19 | 14.96 | 0.27 | 14.92 | 0.53 | |
| 7 | 50 | 49.98 | 0.04 | 49.98 | 0.04 | 49.97 | 0.06 | 49.48 | 1.04 | |
| 8 | 15 | 14.98 | 0.16 | 14.97 | 0.18 | 14.97 | 0.20 | 14.64 | 2.40 | |
| 9 | 5 | 4.99 | 0.20 | 4.99 | 0.28 | 4.94 | 1.20 | 4.73 | 5.40 | |
| Data set 2: μ = 0, σ = 1 | 1 | 150 | 148.93 | 0.71 | 148.79 | 0.81 | 149.14 | 0.57 | 148.712 | 0.86 |
| 2 | 150 | 149.48 | 0.35 | 148.94 | 0.71 | 148.92 | 0.72 | 149.03 | 0.65 | |
| 3 | 50 | 49.16 | 1.68 | 49.22 | 1.56 | 49.24 | 1.52 | 49.31 | 1.38 | |
| 4 | 150 | 149.07 | 0.62 | 148.91 | 0.73 | 148.74 | 0.84 | 148.36 | 1.09 | |
| 5 | 50 | 49.13 | 1.74 | 49.09 | 1.82 | 49.26 | 1.48 | 49.18 | 1.64 | |
| 6 | 15 | 14.43 | 3.80 | 14.20 | 5.33 | 14.61 | 2.60 | 13.975 | 6.83 | |
| 7 | 50 | 48.98 | 2.04 | 49.18 | 1.64 | 48.52 | 2.96 | 49.27 | 1.46 | |
| 8 | 15 | 14.66 | 2.27 | 14.37 | 4.20 | 14.67 | 2.20 | 14.16 | 5.60 | |
| 9 | 5 | 4.72 | 5.60 | 4.49 | 10.20 | 4.48 | 10.40 | 4.37 | 12.60 |
Comparison of estimated longitudinal dispersivity (m) values in various zones using DE, GA, PSO, and CSO
| Data set | Zone | True values (m) | Estimated values using CSO (m) | Absolute % error CSO | Estimated values using PSO (m) | Absolute % error PSO | Estimated values using DE (m) | Absolute % error DE | Estimated values using GA (m) | Absolute % error GA |
|---|---|---|---|---|---|---|---|---|---|---|
| Data set 1 | 1 | 60 | 59.98 | 0.03 | 59.96 | 0.06 | 59.98 | 0.03 | 59.67 | 0.55 |
| 2 | 60 | 59.96 | 0.07 | 59.93 | 0.12 | 59.96 | 0.07 | 59.18 | 1.37 | |
| 3 | 40 | 39.98 | 0.05 | 39.98 | 0.05 | 39.99 | 0.02 | 39.43 | 1.43 | |
| 4 | 60 | 59.99 | 0.02 | 59.96 | 0.07 | 59.98 | 0.03 | 59.50 | 0.83 | |
| 5 | 40 | 39.99 | 0.02 | 39.96 | 0.10 | 39.95 | 0.12 | 39.64 | 0.90 | |
| 6 | 15 | 14.98 | 0.13 | 14.84 | 1.07 | 14.99 | 0.07 | 14.63 | 2.47 | |
| 7 | 40 | 39.99 | 0.03 | 40 | 0.00 | 39.96 | 0.10 | 39.73 | 0.68 | |
| 8 | 15 | 15 | 0.00 | 14.97 | 0.20 | 14.94 | 0.40 | 14.87 | 0.87 | |
| 9 | 10 | 9.97 | 0.30 | 10 | 0.00 | 9.96 | 0.40 | 9.72 | 2.80 | |
| Data set 2: μ = 0, σ = 1 | 1 | 60 | 59.19 | 1.35 | 59.24 | 1.27 | 60.26 | 0.43 | 59.15 | 1.42 |
| 2 | 60 | 59.64 | 0.60 | 59.43 | 0.95 | 59.53 | 0.78 | 59.53 | 0.78 | |
| 3 | 40 | 39.04 | 2.40 | 39.58 | 1.05 | 39.37 | 1.58 | 38.46 | 3.85 | |
| 4 | 60 | 59.72 | 0.47 | 59.29 | 1.18 | 59.79 | 0.35 | 59.06 | 1.57 | |
| 5 | 40 | 39.64 | 0.90 | 39.54 | 1.15 | 40.37 | 0.92 | 39.42 | 1.45 | |
| 6 | 15 | 14.72 | 1.87 | 14.62 | 2.53 | 14.66 | 2.27 | 14.37 | 4.20 | |
| 7 | 40 | 39.68 | 0.80 | 39.57 | 1.08 | 39.61 | 0.98 | 40.36 | 0.90 | |
| 8 | 15 | 14.45 | 3.67 | 14.39 | 4.07 | 14.5 | 3.33 | 14.22 | 5.20 | |
| 9 | 10 | 9.54 | 4.60 | 9.47 | 5.30 | 9.42 | 5.80 | 9.28 | 7.20 |
| Data set | Zone | True values (m) | Estimated values using CSO (m) | Absolute % error CSO | Estimated values using PSO (m) | Absolute % error PSO | Estimated values using DE (m) | Absolute % error DE | Estimated values using GA (m) | Absolute % error GA |
|---|---|---|---|---|---|---|---|---|---|---|
| Data set 1 | 1 | 60 | 59.98 | 0.03 | 59.96 | 0.06 | 59.98 | 0.03 | 59.67 | 0.55 |
| 2 | 60 | 59.96 | 0.07 | 59.93 | 0.12 | 59.96 | 0.07 | 59.18 | 1.37 | |
| 3 | 40 | 39.98 | 0.05 | 39.98 | 0.05 | 39.99 | 0.02 | 39.43 | 1.43 | |
| 4 | 60 | 59.99 | 0.02 | 59.96 | 0.07 | 59.98 | 0.03 | 59.50 | 0.83 | |
| 5 | 40 | 39.99 | 0.02 | 39.96 | 0.10 | 39.95 | 0.12 | 39.64 | 0.90 | |
| 6 | 15 | 14.98 | 0.13 | 14.84 | 1.07 | 14.99 | 0.07 | 14.63 | 2.47 | |
| 7 | 40 | 39.99 | 0.03 | 40 | 0.00 | 39.96 | 0.10 | 39.73 | 0.68 | |
| 8 | 15 | 15 | 0.00 | 14.97 | 0.20 | 14.94 | 0.40 | 14.87 | 0.87 | |
| 9 | 10 | 9.97 | 0.30 | 10 | 0.00 | 9.96 | 0.40 | 9.72 | 2.80 | |
| Data set 2: μ = 0, σ = 1 | 1 | 60 | 59.19 | 1.35 | 59.24 | 1.27 | 60.26 | 0.43 | 59.15 | 1.42 |
| 2 | 60 | 59.64 | 0.60 | 59.43 | 0.95 | 59.53 | 0.78 | 59.53 | 0.78 | |
| 3 | 40 | 39.04 | 2.40 | 39.58 | 1.05 | 39.37 | 1.58 | 38.46 | 3.85 | |
| 4 | 60 | 59.72 | 0.47 | 59.29 | 1.18 | 59.79 | 0.35 | 59.06 | 1.57 | |
| 5 | 40 | 39.64 | 0.90 | 39.54 | 1.15 | 40.37 | 0.92 | 39.42 | 1.45 | |
| 6 | 15 | 14.72 | 1.87 | 14.62 | 2.53 | 14.66 | 2.27 | 14.37 | 4.20 | |
| 7 | 40 | 39.68 | 0.80 | 39.57 | 1.08 | 39.61 | 0.98 | 40.36 | 0.90 | |
| 8 | 15 | 14.45 | 3.67 | 14.39 | 4.07 | 14.5 | 3.33 | 14.22 | 5.20 | |
| 9 | 10 | 9.54 | 4.60 | 9.47 | 5.30 | 9.42 | 5.80 | 9.28 | 7.20 |
Comparison of estimated transverse dispersivity (m) values in various zones using DE, GA, PSO, and CSO
| Data set | Zone | True values (m) | Estimated values using CSO (m) | Absolute % error CSO | Estimated values using PSO (m) | Absolute % error PSO | Estimated values using DE (m) | Absolute % error DE | Estimated values using GA (m)) | Absolute % error GA |
|---|---|---|---|---|---|---|---|---|---|---|
| Data set 1 | 1 | 6 | 5.98 | 0.25 | 5.99 | 0.02 | 5.98 | 0.33 | 5.81 | 3.17 |
| 2 | 6 | 5.91 | 1.50 | 5.99 | 0.03 | 5.96 | 0.67 | 5.72 | 4.67 | |
| 3 | 4 | 3.93 | 1.75 | 4 | 0.00 | 3.94 | 1.50 | 3.82 | 4.50 | |
| 4 | 6 | 5.98 | 0.33 | 6 | 0.00 | 5.99 | 0.17 | 5.92 | 1.33 | |
| 5 | 4 | 3.99 | 0.25 | 3.99 | 0.02 | 3.98 | 0.50 | 3.94 | 1.50 | |
| 6 | 1.5 | 1.48 | 1.33 | 1.5 | 0.00 | 1.49 | 0.67 | 1.46 | 2.67 | |
| 7 | 4 | 4 | 0.00 | 4.01 | 0.25 | 3.98 | 0.50 | 3.91 | 2.25 | |
| 8 | 1.5 | 1.5 | 0.00 | 1.479 | 1.40 | 1.5 | 0.00 | 1.48 | 1.33 | |
| 9 | 1 | 1 | 0.00 | 1 | 0.00 | 0.98 | 2.00 | 0.98 | 2.00 | |
| Data set 2: μ = 0, σ = 1 | 1 | 6 | 5.73 | 4.50 | 5.67 | 5.50 | 5.66 | 5.67 | 5.58 | 7.00 |
| 2 | 6 | 5.61 | 6.50 | 5.58 | 7.00 | 5.7 | 5.00 | 5.59 | 6.83 | |
| 3 | 4 | 3.70 | 7.50 | 3.63 | 9.25 | 3.65 | 8.75 | 3.53 | 11.75 | |
| 4 | 6 | 5.61 | 6.50 | 5.48 | 8.67 | 5.63 | 6.17 | 5.48 | 8.67 | |
| 5 | 4 | 3.77 | 5.75 | 3.57 | 10.75 | 3.69 | 7.75 | 3.5 | 12.50 | |
| 6 | 1.5 | 1.42 | 5.33 | 1.40 | 6.67 | 1.38 | 8.00 | 1.29 | 14.00 | |
| 7 | 4 | 3.84 | 4.00 | 3.68 | 8.00 | 3.77 | 5.75 | 3.62 | 9.50 | |
| 8 | 1.5 | 1.47 | 2.00 | 1.39 | 7.33 | 1.43 | 4.67 | 1.36 | 9.33 | |
| 9 | 1 | 0.95 | 5.00 | 0.93 | 7.00 | 0.96 | 4.00 | 0.93 | 7.00 |
| Data set | Zone | True values (m) | Estimated values using CSO (m) | Absolute % error CSO | Estimated values using PSO (m) | Absolute % error PSO | Estimated values using DE (m) | Absolute % error DE | Estimated values using GA (m)) | Absolute % error GA |
|---|---|---|---|---|---|---|---|---|---|---|
| Data set 1 | 1 | 6 | 5.98 | 0.25 | 5.99 | 0.02 | 5.98 | 0.33 | 5.81 | 3.17 |
| 2 | 6 | 5.91 | 1.50 | 5.99 | 0.03 | 5.96 | 0.67 | 5.72 | 4.67 | |
| 3 | 4 | 3.93 | 1.75 | 4 | 0.00 | 3.94 | 1.50 | 3.82 | 4.50 | |
| 4 | 6 | 5.98 | 0.33 | 6 | 0.00 | 5.99 | 0.17 | 5.92 | 1.33 | |
| 5 | 4 | 3.99 | 0.25 | 3.99 | 0.02 | 3.98 | 0.50 | 3.94 | 1.50 | |
| 6 | 1.5 | 1.48 | 1.33 | 1.5 | 0.00 | 1.49 | 0.67 | 1.46 | 2.67 | |
| 7 | 4 | 4 | 0.00 | 4.01 | 0.25 | 3.98 | 0.50 | 3.91 | 2.25 | |
| 8 | 1.5 | 1.5 | 0.00 | 1.479 | 1.40 | 1.5 | 0.00 | 1.48 | 1.33 | |
| 9 | 1 | 1 | 0.00 | 1 | 0.00 | 0.98 | 2.00 | 0.98 | 2.00 | |
| Data set 2: μ = 0, σ = 1 | 1 | 6 | 5.73 | 4.50 | 5.67 | 5.50 | 5.66 | 5.67 | 5.58 | 7.00 |
| 2 | 6 | 5.61 | 6.50 | 5.58 | 7.00 | 5.7 | 5.00 | 5.59 | 6.83 | |
| 3 | 4 | 3.70 | 7.50 | 3.63 | 9.25 | 3.65 | 8.75 | 3.53 | 11.75 | |
| 4 | 6 | 5.61 | 6.50 | 5.48 | 8.67 | 5.63 | 6.17 | 5.48 | 8.67 | |
| 5 | 4 | 3.77 | 5.75 | 3.57 | 10.75 | 3.69 | 7.75 | 3.5 | 12.50 | |
| 6 | 1.5 | 1.42 | 5.33 | 1.40 | 6.67 | 1.38 | 8.00 | 1.29 | 14.00 | |
| 7 | 4 | 3.84 | 4.00 | 3.68 | 8.00 | 3.77 | 5.75 | 3.62 | 9.50 | |
| 8 | 1.5 | 1.47 | 2.00 | 1.39 | 7.33 | 1.43 | 4.67 | 1.36 | 9.33 | |
| 9 | 1 | 0.95 | 5.00 | 0.93 | 7.00 | 0.96 | 4.00 | 0.93 | 7.00 |
Additionally, to investigate the model's performance with poor quality of observed data, noises are added to hydraulic head and contaminant concentration. Data set 2 is generated by assuming normally distributed error with zero mean and unit variance. The mean percentage error in the estimation of transmissivity, longitudinal dispersivity, and transverse dispersivity values obtained by CSO, PSO, DE, and GA are tabulated in Table 10. It can be seen that for the noisy data set (data set 2), mean percentage error is lowest for CSO. Hence, the finding is that with the noisy data set, estimated aquifer parameters by CSO are closer to the actual values followed by DE, PSO, and GA.
Comparison of mean absolute percentage error for CSO, PSO, DE, and GA
| Data | Without noise | With noise | ||||
|---|---|---|---|---|---|---|
| Mean absolute % error for transmissivity | Mean absolute % error for longitudinal dispersivity | Mean absolute % error for transverse dispersivity | Mean absolute % error for transmissivity | Mean absolute % error for longitudinal dispersivity | Mean absolute % error for transverse dispersivity | |
| CSO | 0.07 | 0.07 | 0.60 | 2.09 | 1.85 | 5.23 |
| PSO | 0.13 | 0.19 | 0.19 | 3.00 | 2.06 | 7.80 |
| DE | 0.22 | 0.14 | 0.70 | 2.59 | 1.85 | 6.19 |
| GA | 1.31 | 1.32 | 2.60 | 3.57 | 2.95 | 9.62 |
| Data | Without noise | With noise | ||||
|---|---|---|---|---|---|---|
| Mean absolute % error for transmissivity | Mean absolute % error for longitudinal dispersivity | Mean absolute % error for transverse dispersivity | Mean absolute % error for transmissivity | Mean absolute % error for longitudinal dispersivity | Mean absolute % error for transverse dispersivity | |
| CSO | 0.07 | 0.07 | 0.60 | 2.09 | 1.85 | 5.23 |
| PSO | 0.13 | 0.19 | 0.19 | 3.00 | 2.06 | 7.80 |
| DE | 0.22 | 0.14 | 0.70 | 2.59 | 1.85 | 6.19 |
| GA | 1.31 | 1.32 | 2.60 | 3.57 | 2.95 | 9.62 |
We have generated bar diagrams to check errors in estimated transmissivity, longitudinal dispersivity, and transverse dispersivity values for data set 2. It can be seen from Figures 7–9 that CSO shows the least error in estimating transmissivity, longitudinal dispersivity, and transverse dispersivity values compared to DE, PSO, and GA.
Comparison of error in estimated transmissivity (m2/day) for noisy data using inverse models.
Comparison of error in estimated transmissivity (m2/day) for noisy data using inverse models.
Comparison of error in estimated longitudinal dispersivity (m) values for noisy data using inverse models.
Comparison of error in estimated longitudinal dispersivity (m) values for noisy data using inverse models.
Comparison of error in estimated transverse dispersivity (m) values for noisy data using inverse models.
Comparison of error in estimated transverse dispersivity (m) values for noisy data using inverse models.
RELIABILITY ANALYSIS
The estimated transmissivity, longitudinal dispersivity, and transverse dispersivity values are tested for reliability analysis. Reliability analysis is ascertained by evaluating the coefficient of variation (CV) and variance-covariance matrix (VCM).
Coefficient of variation
Table 11 shows the CV values, which are estimated for all four models with noisy data (data set 2). Lower CV values of RPCM-CSO indicate better reliability of RPCM-CSO in comparison to RPCM-DE, RPCM-PSO, and RPCM-GA.
Comparison of CV for the estimated transmissivity, longitudinal dispersivity, and transverse dispersivity in various zones for data set 4 using RPCM-GA, RPCM-DE, and RPCM-CSO with the RPCM-PSO model
| Zone no. | Transmissivity | Longitudinal dispersivity | Transverse dispersivity | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CV-RPCM-CSO | CV- RPCM-PSO | CV- RPCM-DE | CV- RPCM GA | CV- RPCM-CSO | CV- RPCM-PSO | CV- RPCM-DE | CV- RPCM-GA | CV- RPCM-CSO | CV- RPCM-PSO | CV- RPCM-DE | CV- RPCM-GA | |
| Zone 1 | 0.0072 | 0.0081 | 0.0058 | 0.0087 | 0.0137 | 0.0128 | 0.0043 | 0.0144 | 0.0471 | 0.0582 | 0.0601 | 0.0753 |
| Zone 2 | 0.0035 | 0.0071 | 0.0073 | 0.0065 | 0.0060 | 0.0096 | 0.0079 | 0.0079 | 0.0695 | 0.0753 | 0.0526 | 0.0733 |
| Zone 3 | 0.0171 | 0.0158 | 0.0154 | 0.0140 | 0.0246 | 0.0106 | 0.0160 | 0.0400 | 0.0811 | 0.1019 | 0.0959 | 0.1331 |
| Zone 4 | 0.0062 | 0.0073 | 0.0085 | 0.0111 | 0.0047 | 0.0120 | 0.0035 | 0.0159 | 0.0695 | 0.0949 | 0.0657 | 0.0949 |
| Zone 5 | 0.0177 | 0.0185 | 0.0150 | 0.0167 | 0.0091 | 0.0116 | 0.0092 | 0.0147 | 0.0610 | 0.1204 | 0.0840 | 0.1429 |
| Zone 6 | 0.0395 | 0.0563 | 0.0267 | 0.0733 | 0.0190 | 0.0260 | 0.0232 | 0.0438 | 0.0563 | 0.0714 | 0.0870 | 0.1628 |
| Zone 7 | 0.0208 | 0.0167 | 0.0305 | 0.0148 | 0.0081 | 0.0109 | 0.0098 | 0.0089 | 0.0417 | 0.0870 | 0.0610 | 0.1050 |
| Zone 8 | 0.0232 | 0.0438 | 0.0225 | 0.0593 | 0.0381 | 0.0424 | 0.0345 | 0.0549 | 0.0204 | 0.0791 | 0.0490 | 0.1029 |
| Zone 9 | 0.0593 | 0.1136 | 0.1161 | 0.1442 | 0.0482 | 0.0559 | 0.0616 | 0.0776 | 0.0526 | 0.0753 | 0.0417 | 0.0753 |
| Zone no. | Transmissivity | Longitudinal dispersivity | Transverse dispersivity | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| CV-RPCM-CSO | CV- RPCM-PSO | CV- RPCM-DE | CV- RPCM GA | CV- RPCM-CSO | CV- RPCM-PSO | CV- RPCM-DE | CV- RPCM-GA | CV- RPCM-CSO | CV- RPCM-PSO | CV- RPCM-DE | CV- RPCM-GA | |
| Zone 1 | 0.0072 | 0.0081 | 0.0058 | 0.0087 | 0.0137 | 0.0128 | 0.0043 | 0.0144 | 0.0471 | 0.0582 | 0.0601 | 0.0753 |
| Zone 2 | 0.0035 | 0.0071 | 0.0073 | 0.0065 | 0.0060 | 0.0096 | 0.0079 | 0.0079 | 0.0695 | 0.0753 | 0.0526 | 0.0733 |
| Zone 3 | 0.0171 | 0.0158 | 0.0154 | 0.0140 | 0.0246 | 0.0106 | 0.0160 | 0.0400 | 0.0811 | 0.1019 | 0.0959 | 0.1331 |
| Zone 4 | 0.0062 | 0.0073 | 0.0085 | 0.0111 | 0.0047 | 0.0120 | 0.0035 | 0.0159 | 0.0695 | 0.0949 | 0.0657 | 0.0949 |
| Zone 5 | 0.0177 | 0.0185 | 0.0150 | 0.0167 | 0.0091 | 0.0116 | 0.0092 | 0.0147 | 0.0610 | 0.1204 | 0.0840 | 0.1429 |
| Zone 6 | 0.0395 | 0.0563 | 0.0267 | 0.0733 | 0.0190 | 0.0260 | 0.0232 | 0.0438 | 0.0563 | 0.0714 | 0.0870 | 0.1628 |
| Zone 7 | 0.0208 | 0.0167 | 0.0305 | 0.0148 | 0.0081 | 0.0109 | 0.0098 | 0.0089 | 0.0417 | 0.0870 | 0.0610 | 0.1050 |
| Zone 8 | 0.0232 | 0.0438 | 0.0225 | 0.0593 | 0.0381 | 0.0424 | 0.0345 | 0.0549 | 0.0204 | 0.0791 | 0.0490 | 0.1029 |
| Zone 9 | 0.0593 | 0.1136 | 0.1161 | 0.1442 | 0.0482 | 0.0559 | 0.0616 | 0.0776 | 0.0526 | 0.0753 | 0.0417 | 0.0753 |
Variance-covariance matrix
VCM for noisy data using the RPCM-GA model for the estimated transverse dispersivity in various zones
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.1764 | 0.1722 | 0.1974 | 0.2184 | 0.21 | 0.0882 | 0.1596 | 0.0588 | 0.0294 |
| 2 | 0.1722 | 0.1681 | 0.1927 | 0.2132 | 0.205 | 0.0861 | 0.1558 | 0.0574 | 0.0287 |
| 3 | 0.1974 | 0.1927 | 0.2209 | 0.2444 | 0.235 | 0.0987 | 0.1786 | 0.0658 | 0.0329 |
| 4 | 0.2184 | 0.2132 | 0.2444 | 0.2704 | 0.26 | 0.1092 | 0.1976 | 0.0728 | 0.0364 |
| 5 | 0.21 | 0.205 | 0.235 | 0.26 | 0.25 | 0.105 | 0.19 | 0.07 | 0.035 |
| 6 | 0.0882 | 0.0861 | 0.0987 | 0.1092 | 0.105 | 0.0441 | 0.0798 | 0.0294 | 0.0147 |
| 7 | 0.1596 | 0.1558 | 0.1786 | 0.1976 | 0.19 | 0.0798 | 0.1444 | 0.0532 | 0.0266 |
| 8 | 0.0588 | 0.0574 | 0.0658 | 0.0728 | 0.07 | 0.0294 | 0.0532 | 0.0196 | 0.0098 |
| 9 | 0.0294 | 0.0287 | 0.0329 | 0.0364 | 0.035 | 0.0147 | 0.0266 | 0.0098 | 0.0049 |
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.1764 | 0.1722 | 0.1974 | 0.2184 | 0.21 | 0.0882 | 0.1596 | 0.0588 | 0.0294 |
| 2 | 0.1722 | 0.1681 | 0.1927 | 0.2132 | 0.205 | 0.0861 | 0.1558 | 0.0574 | 0.0287 |
| 3 | 0.1974 | 0.1927 | 0.2209 | 0.2444 | 0.235 | 0.0987 | 0.1786 | 0.0658 | 0.0329 |
| 4 | 0.2184 | 0.2132 | 0.2444 | 0.2704 | 0.26 | 0.1092 | 0.1976 | 0.0728 | 0.0364 |
| 5 | 0.21 | 0.205 | 0.235 | 0.26 | 0.25 | 0.105 | 0.19 | 0.07 | 0.035 |
| 6 | 0.0882 | 0.0861 | 0.0987 | 0.1092 | 0.105 | 0.0441 | 0.0798 | 0.0294 | 0.0147 |
| 7 | 0.1596 | 0.1558 | 0.1786 | 0.1976 | 0.19 | 0.0798 | 0.1444 | 0.0532 | 0.0266 |
| 8 | 0.0588 | 0.0574 | 0.0658 | 0.0728 | 0.07 | 0.0294 | 0.0532 | 0.0196 | 0.0098 |
| 9 | 0.0294 | 0.0287 | 0.0329 | 0.0364 | 0.035 | 0.0147 | 0.0266 | 0.0098 | 0.0049 |
VCM for noisy data using the RPCM-DE model for the estimated transverse dispersivity in various zones
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.1156 | 0.102 | 0.119 | 0.1258 | 0.1054 | 0.0408 | 0.0782 | 0.0238 | 0.0136 |
| 2 | 0.102 | 0.09 | 0.105 | 0.111 | 0.093 | 0.036 | 0.069 | 0.021 | 0.012 |
| 3 | 0.119 | 0.105 | 0.1225 | 0.1295 | 0.1085 | 0.042 | 0.0805 | 0.0245 | 0.014 |
| 4 | 0.1258 | 0.111 | 0.1295 | 0.1369 | 0.1147 | 0.0444 | 0.0851 | 0.0259 | 0.0148 |
| 5 | 0.1054 | 0.093 | 0.1085 | 0.1147 | 0.0961 | 0.0372 | 0.0713 | 0.0217 | 0.0124 |
| 6 | 0.0408 | 0.036 | 0.042 | 0.0444 | 0.0372 | 0.0144 | 0.0276 | 0.0084 | 0.0048 |
| 7 | 0.0782 | 0.069 | 0.0805 | 0.0851 | 0.0713 | 0.0276 | 0.0529 | 0.0161 | 0.0092 |
| 8 | 0.0238 | 0.021 | 0.0245 | 0.0259 | 0.0217 | 0.0084 | 0.0161 | 0.0049 | 0.0028 |
| 9 | 0.0136 | 0.012 | 0.014 | 0.0148 | 0.0124 | 0.0048 | 0.0092 | 0.0028 | 0.0016 |
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.1156 | 0.102 | 0.119 | 0.1258 | 0.1054 | 0.0408 | 0.0782 | 0.0238 | 0.0136 |
| 2 | 0.102 | 0.09 | 0.105 | 0.111 | 0.093 | 0.036 | 0.069 | 0.021 | 0.012 |
| 3 | 0.119 | 0.105 | 0.1225 | 0.1295 | 0.1085 | 0.042 | 0.0805 | 0.0245 | 0.014 |
| 4 | 0.1258 | 0.111 | 0.1295 | 0.1369 | 0.1147 | 0.0444 | 0.0851 | 0.0259 | 0.0148 |
| 5 | 0.1054 | 0.093 | 0.1085 | 0.1147 | 0.0961 | 0.0372 | 0.0713 | 0.0217 | 0.0124 |
| 6 | 0.0408 | 0.036 | 0.042 | 0.0444 | 0.0372 | 0.0144 | 0.0276 | 0.0084 | 0.0048 |
| 7 | 0.0782 | 0.069 | 0.0805 | 0.0851 | 0.0713 | 0.0276 | 0.0529 | 0.0161 | 0.0092 |
| 8 | 0.0238 | 0.021 | 0.0245 | 0.0259 | 0.0217 | 0.0084 | 0.0161 | 0.0049 | 0.0028 |
| 9 | 0.0136 | 0.012 | 0.014 | 0.0148 | 0.0124 | 0.0048 | 0.0092 | 0.0028 | 0.0016 |
VCM for noisy data using the RPCM-CSO model for the estimated transverse dispersivity in various zones
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.09 | 0.105 | 0.096 | 0.135 | 0.093 | 0.039 | 0.084 | 0.018 | 0.021 |
| 2 | 0.105 | 0.1225 | 0.112 | 0.1575 | 0.1085 | 0.0455 | 0.098 | 0.021 | 0.0245 |
| 3 | 0.096 | 0.112 | 0.1024 | 0.144 | 0.0992 | 0.0416 | 0.0896 | 0.0192 | 0.0224 |
| 4 | 0.135 | 0.1575 | 0.144 | 0.2025 | 0.1395 | 0.0585 | 0.126 | 0.027 | 0.0315 |
| 5 | 0.093 | 0.1085 | 0.0992 | 0.1395 | 0.0961 | 0.0403 | 0.0868 | 0.0186 | 0.0217 |
| 6 | 0.039 | 0.0455 | 0.0416 | 0.0585 | 0.0403 | 0.0169 | 0.0364 | 0.0078 | 0.0091 |
| 7 | 0.084 | 0.098 | 0.0896 | 0.126 | 0.0868 | 0.0364 | 0.0784 | 0.0168 | 0.0196 |
| 8 | 0.018 | 0.021 | 0.0192 | 0.027 | 0.0186 | 0.0078 | 0.0168 | 0.0036 | 0.0042 |
| 9 | 0.021 | 0.0245 | 0.0224 | 0.0315 | 0.0217 | 0.0091 | 0.0196 | 0.0042 | 0.0049 |
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.09 | 0.105 | 0.096 | 0.135 | 0.093 | 0.039 | 0.084 | 0.018 | 0.021 |
| 2 | 0.105 | 0.1225 | 0.112 | 0.1575 | 0.1085 | 0.0455 | 0.098 | 0.021 | 0.0245 |
| 3 | 0.096 | 0.112 | 0.1024 | 0.144 | 0.0992 | 0.0416 | 0.0896 | 0.0192 | 0.0224 |
| 4 | 0.135 | 0.1575 | 0.144 | 0.2025 | 0.1395 | 0.0585 | 0.126 | 0.027 | 0.0315 |
| 5 | 0.093 | 0.1085 | 0.0992 | 0.1395 | 0.0961 | 0.0403 | 0.0868 | 0.0186 | 0.0217 |
| 6 | 0.039 | 0.0455 | 0.0416 | 0.0585 | 0.0403 | 0.0169 | 0.0364 | 0.0078 | 0.0091 |
| 7 | 0.084 | 0.098 | 0.0896 | 0.126 | 0.0868 | 0.0364 | 0.0784 | 0.0168 | 0.0196 |
| 8 | 0.018 | 0.021 | 0.0192 | 0.027 | 0.0186 | 0.0078 | 0.0168 | 0.0036 | 0.0042 |
| 9 | 0.021 | 0.0245 | 0.0224 | 0.0315 | 0.0217 | 0.0091 | 0.0196 | 0.0042 | 0.0049 |
VCM for noisy data using the RPCM-PSO model for the estimated transverse dispersivity in various zones
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.1563 | 0.18235 | 0.16672 | 0.23445 | 0.16151 | 0.06773 | 0.14588 | 0.03126 | 0.03647 |
| 2 | 0.1461 | 0.17045 | 0.15584 | 0.21915 | 0.15097 | 0.06331 | 0.13636 | 0.02922 | 0.03409 |
| 3 | 0.1116 | 0.1302 | 0.11904 | 0.1674 | 0.11532 | 0.04836 | 0.10416 | 0.02232 | 0.02604 |
| 4 | 0.1743 | 0.20335 | 0.18592 | 0.26145 | 0.18011 | 0.07553 | 0.16268 | 0.03486 | 0.04067 |
| 5 | 0.1296 | 0.1512 | 0.13824 | 0.1944 | 0.13392 | 0.05616 | 0.12096 | 0.02592 | 0.03024 |
| 6 | 0.0543 | 0.06335 | 0.05792 | 0.08145 | 0.05611 | 0.02353 | 0.05068 | 0.01086 | 0.01267 |
| 7 | 0.1206 | 0.1407 | 0.12864 | 0.1809 | 0.12462 | 0.05226 | 0.11256 | 0.02412 | 0.02814 |
| 8 | 0.0321 | 0.03745 | 0.03424 | 0.04815 | 0.03317 | 0.01391 | 0.02996 | 0.00642 | 0.00749 |
| 9 | 0.027 | 0.0315 | 0.0288 | 0.0405 | 0.0279 | 0.0117 | 0.0252 | 0.0054 | 0.0063 |
| 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | |
|---|---|---|---|---|---|---|---|---|---|
| 1 | 0.1563 | 0.18235 | 0.16672 | 0.23445 | 0.16151 | 0.06773 | 0.14588 | 0.03126 | 0.03647 |
| 2 | 0.1461 | 0.17045 | 0.15584 | 0.21915 | 0.15097 | 0.06331 | 0.13636 | 0.02922 | 0.03409 |
| 3 | 0.1116 | 0.1302 | 0.11904 | 0.1674 | 0.11532 | 0.04836 | 0.10416 | 0.02232 | 0.02604 |
| 4 | 0.1743 | 0.20335 | 0.18592 | 0.26145 | 0.18011 | 0.07553 | 0.16268 | 0.03486 | 0.04067 |
| 5 | 0.1296 | 0.1512 | 0.13824 | 0.1944 | 0.13392 | 0.05616 | 0.12096 | 0.02592 | 0.03024 |
| 6 | 0.0543 | 0.06335 | 0.05792 | 0.08145 | 0.05611 | 0.02353 | 0.05068 | 0.01086 | 0.01267 |
| 7 | 0.1206 | 0.1407 | 0.12864 | 0.1809 | 0.12462 | 0.05226 | 0.11256 | 0.02412 | 0.02814 |
| 8 | 0.0321 | 0.03745 | 0.03424 | 0.04815 | 0.03317 | 0.01391 | 0.02996 | 0.00642 | 0.00749 |
| 9 | 0.027 | 0.0315 | 0.0288 | 0.0405 | 0.0279 | 0.0117 | 0.0252 | 0.0054 | 0.0063 |
DISCUSSION
This study is an attempt to check the effectiveness of four important optimization algorithms (GA, DE, CSO, PSO) in estimating aquifer parameters (transmissivity, longitudinal dispersivity, and transverse dispersivity) of a hypothetical confined aquifer. It has been observed that, for non-noisy observed data (hydraulic head and contaminant concentration), DE, CSO, and PSO are more accurate in estimating aquifer parameters than GA. However, for noisy observed data, CSO outperforms the DE, PSO, and GA. This is an important finding, in the sense that in real field scenarios, the observed data are usually erroneous, and in such cases use of CSO for inverse modeling will be a better choice than GA, DE, and PSO. However, the computational time of CSO-based optimization models is greater compared to DE and PSO. In the seeking mode of CSO, each cat creates its multiple copies in the search space. Hence, in CSO for every iteration, the number of fitness function evolutions are greater in comparison with other optimization algorithms. However, the computational speed of CSO can be drastically reduced by using high-performance computing systems, a graphics processing unit, vectorizing the codes, parallel computing, and using surrogate simulators. The RPCM model can be used to train surrogate models such as artificial neural network, deep learning, radial basis function, support vector machine, kriging, etc. The SO model developed by coupling a surrogate model with CSO will be computationally more efficient than RPCM-CSO.
CONCLUSIONS
In this study, a mesh-free RPCM model is developed to simulate the groundwater flow and contaminant transport processes of a confined aquifer. The hydraulic head and contaminant concentration obtained by the RPCM model are found to be in good agreement with the solution of FEM. Further, four optimization models are developed based on GA, DE, PSO, and CSO. By coupling the RPCM model separately with GA, DE, PSO, and CSO, the S-O models are developed. The S-O models are used to compute transmissivity, longitudinal dispersivity, and transverse dispersivity by minimizing a composite objective function. The composite objective is a summation of sum of the weighted squared differences of observed and simulated hydraulic head and sum of the weighted squared differences of observed and simulated contaminant concentrations at the monitoring wells. The S-O model is applied to compute aquifer parameters of a hypothetical confined aquifer with nine zones. It is observed that RPCM-CSO, RPCM-DE, and RPCM-PSO models are more accurate in estimating aquifer parameters than RPCM-GA. The performance of RPCM-CSO and RPCM-DE are comparable and it is difficult to ascertain the superiority of one over the other.
Further, high measurement errors (noises) are introduced to the observed hydraulic head and contaminant concentrations' values to check the performance of the S-O model in estimating aquifer parameters when observed data sets are noisy. Here, it is observed that the average percentage errors in estimating aquifer parameters by the RPCM-CSO model are less than those obtained by RPCM-GA, RPCM-PSO, and RPCM-DE. For the noisy observed data set, the reliability analysis also suggests the superiority of the RPCM-CSO model over RPCM-GA, RPCM-DE, and RPCM-PSO. It is evident from the results that CSO is more accurate in estimating aquifer parameters than GA, DE, and PSO with the noisy observed data set. The RPCM-CSO model can be effectively used for estimating aquifer parameters or contaminant source identification using inverse modeling for aquifers with more complex hydrogeological features.
%20cropped.png?versionId=7142)