Groundwater resources are considered among the most vital and cost-effective sources of freshwater in arid and semi-arid climates. Excessive extraction from these resources has led to a severe drop in the water table and adverse environmental consequences. The present study was conducted to apply and evaluate the efficiency of neuro-fuzzy inference systems at a local scale to accurately predict groundwater level fluctuations in the Zaveh-Torbat-e Heydarieh aquifer with an approach toward the optimal management of water resources.

In this research, monthly groundwater level data from 18 observation wells (piezometers) over a 20-year period (1996–2016) were utilized. For this purpose, data-driven models including Multilayer Perceptron (MLP) neural network, Radial Basis Function (RBF) network, and Adaptive Neuro-Fuzzy Inference System (ANFIS) were developed. To guarantee model robustness and prevent overfitting, the $k$-fold cross-validation method (with $k=5$) was employed. The input variables of the model included climatic parameters of evaporation, maximum and minimum temperatures, total rainfall, average relative humidity, time step (month), and time-lagged groundwater level (previous month), while the current water level was defined as the output variable.

The research findings indicated that all three intelligent models possess acceptable accuracy in predicting fluctuations; however, the neuro-fuzzy algorithm (ANFIS) demonstrated highly superior performance in estimating the monthly groundwater level. (RMSE) values for the MLP, RBF, and ANFIS models were obtained as 0.026, 0.062, and 0.003 m, respectively.

The results of this study confirm the high efficiency of hybrid artificial intelligence models, particularly the ANFIS system, in analyzing complex and non-linear hydrological systems.