Groundwater resources are becoming increasingly important, especially in arid and semi-arid regions affected by climate change [1
]. In many areas, surface water resources are rapidly decreasing, inducing larger pressure on groundwater [2
]. Thus, the prediction of changes in groundwater level (GWL) is becoming increasingly essential for sustainable use [3
]. Reliable prediction of GWL, however, requires extensive and labor-consuming observations. They involve climatic, hydrological, geological variables, and land use change. Furthermore, some of these variables (e.g., climate and land use change) change over time, which increases the model complexity [6
]. Therefore, depending on the available information and uncertainties, different models have been developed to simulate the behavior of GWL changes. In general, three main model types are used to simulate and predict GWL: physical, numerical, and regression models.
Physical models are suitable for prediction but suffer from high cost of construction and require detailed physical information of the aquifer. Numerical models do not have the limitations of physical models but require information on the aquifer’s geology, such as hydraulic conductivity, storage coefficient, and aquifer thickness. Obtaining such information is still difficult, and in many areas, especially in Iran, it is either not available or associated with significant errors. As a result, numerical models are associated with weak performance. Some of these models are PMWIN, FFLOW, GMS, and Visual MODFLOW [8
]. Regression models do not have the limitations of physical models and do not require the information needed by numerical models. They require time-dependent observations that affect GWL (e.g., precipitation, temperature, and aquifer withdrawal) [11
]. However, these models have shown poor performance in complex situations, especially when the number of independent variables is high [11
Machine learning (ML) and artificial intelligence (AI) models are sophisticated regression methods that compensate for the shortcomings of simple regression models. ML and AI models have been used in various disciplines showing a good performance compared to other regression models [13
]. Among them, artificial neural networks (ANN), support vector machines (SVM), Bayesian network (BN), fuzzy inference system (FIS), and adaptive neuro-fuzzy inference systems (ANFIS) have attracted recent attention [18
]. These models do not require structural and physical information of the system, instead time-dependent variables are considered. They have been used extensively by various researchers in groundwater management and GWL prediction [1
ANFIS is a model developed from the combination of ANN and FIS [22
]. This model has shown promising performance in many research fields [19
]. In GWL simulations, ANFIS has been used by several researchers (Table 1
). From 2009 to 2020, researchers compared the performance of ANFIS in predicting GWL with other machine learning models. In general, the capability of this model has revealed efficient performance in groundwater simulation. However, in some cases, it showed poor performance when trapped in local minima [26
]. To improve the modeling quality, metaheuristic algorithms have been developed in the training of ANFIS [27
]. These algorithms can use large amounts of data and are usually not trapped in local minima and show a good convergence rate.
According to Table 1
, during the last ten years, only one study has used evolutionary algorithms to improve the performance of ANFIS. However, it is necessary to examine other algorithms to improve the performance of ANFIS. For example, Seifi et al. (2019) predicted GWL with lagged GWL values [28
]. Variables such as temperature, precipitation, evaporation, and aquifer withdrawal, in addition to lagged GWL, have been used as input. Few studies have considered all these input variables simultaneously.
Summary of application of ANFIS models in GWL prediction.
Summary of application of ANFIS models in GWL prediction.
|No||Reference||Models Used||Input Variables *|
|2||||ANFIS, ANN||GWL, T, P|
|3||||ANFIS, ANN||GWL, P, E, T, H|
|6||||GP, ANN, ANFIS, SVM||GWL, P, E|
|7||||GP, ANFIS||GWL, P, E|
|8||||Wavelet-ANFIS, Wavelet, ANFIS, ANN||GWL, P, E, average Q|
|9||||ANFIS, SVR||R, E, Q, W|
|10||||ANFIS, SVM, ANN||GWL, SWL, P, T|
|12||||Optimization of ANN, ANFIS, SVM, with GOA, PSO, WA, CSO, and KA||GWL|
|13||This study||Optimization of ANFIS with GA, PSO, ACOR, and DE||GWL, P, T, R, Q, E|
In recent years, several algorithms such as genetic algorithm (GA), particle swarm optimization (PSO), ant colony optimization for continuous domains (ACOR), and differential evolution (DE) have been used for better training of ANFIS. The results of these hybrid models are more accurate than that of the ANFIS itself. In a study, Azad et al. (2018) used GA, PSO, ACOR, and DE algorithms in ANFIS training to simulate the water quality of the Gorganrood River in Iran [40
]. They showed that DE has the highest accuracy compared to other evolutionary algorithms in river quality simulation. In another study, they used the algorithms to model rainfall-runoff in Isfahan [41
]. They reported that the ACOR algorithm provided the best accuracy among the investigated models. Yang et al. (2019) used ANFIS-GA and ANFIS-PSO hybrid models to predict landslides. The proposed models were capable of statistically predicting the landslides with an excellent level of accuracy [42
Arya Azar et al. (2021) used ANFIS to predict the longitudinal dispersion coefficient of rivers. The Harris hawks optimization (HHO) algorithm was used to increase the model’s performance, and results were compared with experimental models and LSSVM in predicting the longitudinal dispersion coefficient of the river [43
]. The results showed that the HHO-ANFIS hybrid model had a higher performance compared to other models. Ghordoyee Milan et al. (2021) used PSO, gray wolf optimization (GWO), and HHO algorithms to improve the results of ANFIS for optimal groundwater withdrawal. Results showed that HHO had the highest performance in improving the ANFIS model [19
]. In another study, Kayhomayoon et al. (2021) evaluated the efficiency of ANFIS, ANFIS-HHO, and LSSVM models in predicting the shortage of groundwater reserves [1
]. They reported that the optimized ANFIS with HHO had higher performance than other models.
Considering the importance of GWL prediction with the least possible information and a variety of machine learning models, along with the ability of metaheuristic algorithms to find optimal global solutions, GWL prediction was investigated in the present study using ANFIS. In other words, the ANFIS model sometimes does not correctly identify the behavior of the global minimum points and the model error increases due to the lack of proper prediction in these points [28
]. Therefore, to improve the performance of the ANFIS model, it is necessary to use evolutionary algorithms that train the ANFIS model well. However, the main purpose of this study is to evaluate the efficiency of different ANFIS-metaheuristic hybrid models for GWL prediction in the Lake Urmia watershed. The performance of these models should be compared and, finally, the best model can be selected for forecasting.
The ANFIS was used together with several algorithms to find the best hybrid model in GWL prediction. For this purpose, GA, PSO, ACOR, and DE were used to improve the prediction performance of the ANFIS model. Since ANFIS-metaheuristic hybrid models have rarely been used in predicting GWL, investigating the performance of hybrid models with different algorithms is innovative. Since observation wells are close to the river, river flow can affect the GWL. Thus, river flow for wells close to rivers along with temperature, precipitation, and aquifer withdrawal were used to predict GWL. In total, 11 input patterns from various combinations of variables were used to develop ANFIS hybrid models to predict GWL. The results of the models, along with the input patterns, are provided using error evaluation criteria and graphs, and, finally, the most appropriate model and input pattern are proposed for GWL prediction.
The ANFIS model was used together with the metaheuristic algorithms to improve its training performance in predicting the GWL of the two observation wells. Table 5
shows the parameters and specifications of the ANFIS and hybrid models, namely, ANFIS-GA, ANFIS-PSO, ANFIS-ACOR, and ANFIS-DE. The maximum number of iterations varied for each model, which was determined based on trial and error, and better results were not obtained for more iterations. To obtain appropriate values for each model, data were given in different ranges as initial values, and the best value for each parameter was determined. The Sugeno-type function was best for the ANFIS structure. In total, 10 rules were selected to predict GWL (8–12 were tested). Linear output function was selected as the best output function for ANFIS. The initial population for GA was selected to be 100 as the most suitable from the tested 50, 75, 100, and 125 populations. The value 0.7 was selected as the most appropriate for the mutation percentage; from a range between 0.2 and 0.6, the most appropriate value was 0.3 for the crossover percentage. After considering different ranges for other algorithms, the optimal values for the parameters of each model were selected according to Table 5
. GA and PSO have the highest number of parameters, and ACOR and DE the lowest to achieve the optimal result. For example, except for the initial population selection, the ACOR algorithm can be adjusted with only two parameters of deviation distance rate and selection pressure (Table 5
3.1. Observation Well P1
The performance of the input patterns in the prediction of GWL is shown in Table 6
. For the ANFIS model, there was a significant inconsistency for all input patterns. Therefore, the use of metaheuristic optimization algorithms was used to improve the ANFIS model. Input pattern L, in which all input variables were used, was the best input pattern. The ANFIS-PSO hybrid model showed best performance with MAPE, NSE, and RMSE criteria equal to 0.00019, 0.95, and 0.28 m for the test data, respectively. However, the DE and PSO algorithms were very close to each other. Regardless of the input pattern, the performance of the PSO algorithm in training the ANFIS model was the best compared to other algorithms. It can be concluded that the most suitable metaheuristic algorithms were PSO, DE, and GA, respectively, while the ACOR algorithm exerted the weakest prediction accuracy for most input patterns.
The lowest performance was for input pattern E (temperature and evaporation). In general, the input patterns that included only two input variables had low simulation accuracy. Thus, few input variables are not capable of correctly identifying the nonlinear relationships. In contrast, input patterns with three or more variables showed good accuracy. This suggests that the participation of different input variables can correctly distinguish nonlinear relationships between the input variables and GWL. Input patterns that included lagged GWL resulted in higher prediction performance, which reveals its importance for prediction. Input patterns D, E, and G did not contain lagged GWL (previous month) as input, and, therefore, the performance of the models using these input patterns significantly decreased. For example, input pattern F, which included lagged GWL for the previous month compared to input pattern D, increased the prediction accuracy remarkably. The only difference between input patterns G and E was the presence of groundwater withdrawal. The presence of this input variable did not significantly improve the GWL prediction. Thus, it is necessary to consider input variables that have both proportional and inverse relations with the output variable. In other words, an input pattern that includes both increasing and decreasing effects on the output variable results in higher performance. In input patterns F, J, and H, temperature, evaporation, and precipitation were added to the input patterns, respectively, along with GWL and groundwater withdrawal. The results showed that evaporation, temperature, and precipitation are the most crucial input variables, respectively. For example, precipitation is essential in aquifer recharge and directly affects the GWL, while evaporation has indirect effects on GWL. As a result, the ANFIS-DE model resulted in the highest accuracy for input pattern J with MAPE, NSE, and RMSE values equal to 0.00027 m, 0.89, and 0.44 m, respectively, for the test data. Finally, the GWL for the previous month, evaporation, and temperature, were variables that improved the prediction accuracy of the models.
One of the best methods to interpret RMSE values and understand if they are acceptable for model evaluation is to evaluate them with standard deviation values (STD). It is known from previous studies that RMSE values less than half of the SD of the measured data might be considered low and acceptable [55
]. Hence, considering the RMSE and STD values for both training and test data, the RMSE value is less than half the STD value in most scenarios. This indicates that the RMSE value for the second observation well is within the acceptable range.
Observed and predicted test data are depicted in Figure 6
for observation well P1. The results of ANFIS show that, in some months, there is a large prediction error, which can be seen for time steps 13, 37, and 49. However, improved ANFIS shows acceptable results. This be seen in the scatter plots of the hybrid models with R2
values higher than 0.9.
Taylor’s diagram of the selected input pattern of each model is depicted in Figure 7
. The x and y axes indicate the standard deviation of the data. The quarter-circle arc shows the correlation coefficient of arbitrary and observation data, which varies from 0 to 1. The observation data lie on the x-axis, and predictions close to the x-axis indicate a strong correlation with observations. The green arcs indicate root mean square deviation (RMSD). The highest correlation with observations (>97%) was obtained using the ANFIS-PSO model. However, in other models, the correlation was lower than 95%. The standard deviation for the ANFIS-PSO model is equal to that of the observations. Results depicted in Figure 8
, along with the previously obtained results, indicate that the ANFIS-PSO hybrid model leads to the best prediction performance compared to the other models.
The time series prediction for the entire studied period is shown in Figure 8
. The GWL has a sinusoidal trend because a large amount of groundwater is exploited in six months of the year. During the rest of the year, due to the lack of agricultural irrigation and reduced aquifer withdrawal, a rise in GWL can be observed. Since the temperature and evaporation are higher in the first half of the year, the pattern of changes in GWL, temperature, and evaporation are similar, and, therefore, the temperature and evaporation are considered as the effective variables in GWL prediction. However, given the changes in the trend of GWL throughout the study period, hybrid models gave accurate predictions. The simulation results show that for well P1 (observation well with no information on river flow), GWL dropped more than 2 m, i.e., about 0.12 m/year on average.
3.2. Observation Well P2
Observation well P2 was selected so that along with the effects of meteorological variables, the effects of the presence or absence of river flow can be investigated on GWL prediction. Therefore, the input patterns considered for this well included the river flow in addition to the same variables for well P1. The results of the evaluation criteria for the prediction of training and testing data for the ANFIS and hybrid models are given in Table 7
. Unlike the results obtained from well P1, different input patterns were selected as the appropriate solution for each model. Input pattern L showed better performance for the ANFIS model. This input pattern included lagged GWL, groundwater withdrawal, river flow, precipitation, and temperature. Using this input pattern, MAPE, NSE, and RMSE were obtained equal to 0.00055 m, 0.86, and 0.79 m, respectively. Input pattern G was selected for the ANFIS-GA and ANFIS-DE hybrid models. Of the two, the DE algorithm performed better than the GA. The inputs of this input pattern included lagged GWL, evaporation, precipitation, groundwater withdrawal, and river flow. The only difference between this input pattern and the selected input pattern in the ANFIS (input pattern L) was evaporation instead of temperature. The values of MAPE, NSE, and RMSE for the test data were obtained: 0.00042 m, 0.95 , and 0.61 m for GA, respectively, as well as 0.00034 m, 0.96, and 0.53 m for DE. The input patterns N and Q were the best for the ANFIS-PSO and ANFIS-ACOR hybrid models, respectively. Input pattern N included all input variables except precipitation, while input pattern Q included all the input variables. Using the ANFIS-ACOR hybrid model, input pattern Q resulted in MAPE, NSE, and RMSE evaluation criteria equal to 0.0003 m, 0.97, and 0.45 m, respectively. The results for the observation well P2 show that for proper GWL simulation, it is necessary to consider input patterns with more than three variables. None of the input patterns with lower than four input variables resulted in proper accuracy in GWL prediction, which indicates the complex nonlinear relationships between the inputs and the output. As with the results obtained for P1, the GWL at the previous month had the highest impact on GWL prediction. In input pattern G, which was selected for the ANFIS-GA and ANFIS-DE hybrid models, the river flow was not used as input, but for other models, river flow was effective in improving the prediction accuracy.
Input patterns E, F, J, and K did not include lagged GWL and they had the weakest prediction performance. Input pattern A, with only lagged GWL, provided a better result than input patterns E, F, J, and K.
Input patterns B and C included lagged GWL and aquifer withdrawal and either river flow or precipitation. The performance was better for input pattern B compared to input pattern C. It can be inferred that the river flow is more important than precipitation as an input. However, both variables bring similar information to the models. The best input pattern for the ANFIS-PSO model was input pattern N. Comparing the RMSE and STD values in most scenarios, especially in the selected scenario of each model, shows that RMSE values are always less than half of the STD values, which are within the acceptable range [55
The time series for observations and predictions are shown in Figure 9
. Similar to the results of observation well P1, the scatter plot of the ANFIS model indicated less accuracy (R2
= 0.71), while the hybrid models had an acceptable prediction performance, especially the ANFIS-ACOR hybrid model (R2
= 0.97). Therefore, input pattern Q, which includes all variables, is capable of predicting the GWL in the observation wells near the rivers with appropriate accuracy using the ANFIS-ACOR hybrid model.
Taylor’s diagram for the selected input patterns for observation well P2 is shown in Figure 10
. For the ANFIS-ACOR model, the correlation coefficient is about 0.98, which indicates efficient prediction. The standard deviation for both observations and predictions obtained by this model was similar and equal to 2.6 m. The correlation coefficient, standard deviation, and RMSD for other models were similar. The standard deviation of the ANFIS-PSO was close to that of the observation data.
The time input pattern prediction for the whole time period for all models is shown in Figure 11
. Similar to the results obtained for observation well P1, hybrid models were able to recognize the trend in GWL. However, the ANFIS model resulted in high prediction errors for three time steps. As shown in the figure, these were December 2005, February 2011, and April 2014. Therefore, all models, except the ANFIS, are capable of a reliable GWL prediction in the study period. The simulation results show that for well P2 (observation well with river flow as an independent input variable), GWL dropped more than 8 m, i.e., ca. 0.57 m/year on average. Therefore, the annual drop in this area is more than that of the area of the observation well P1.
Given the importance of GWL in assessing the quantitative status of aquifers for decision-making problems, from the perspective of water resources managers, it is essential to develop predictive models to investigate the status of groundwater resources for this purpose, ANFIS and hybrid ANFIS-metaheuristic algorithms were studied to simulate GWL in an aquifer. Several input variables, including the GWL at the previous month, precipitation, temperature, evaporation, and withdrawal, were considered using experimental input patterns. This approach was performed for two observation wells. For the observation well P1, an input pattern in which all input variables were used gave the best results using the ANFIS-PSO hybrid model with MAPE, RMSE, and NSE values equal to 0.00019, 0.95, and 0.28 m for the test data, respectively. For the observation well P2, river flow was added to the input patterns where the ANFIS-ACOR model showed the best performance with MAPE, NSE, and RMSE of 0.0003, 0.97, and 0.45 m for test data, respectively.
None of the input patterns with less than four input variables, showed acceptable performance in predicting the GWL. Results also showed that river flow generally can increase the prediction accuracy. Finally, results showed appropriate performance of the hybrid models for GWL prediction. This approach can be used in other areas with limited input data to predict GWL. The use of metaheuristic algorithms proposed in this study increased the prediction performance. Considering the uncertainty in input variables, evaluating new algorithms to improve the results, and investigating the effects of climate change on GWL are among the research topics that are suggested for future investigations.