Intelligent approach to predict future groundwater level based on artificial neural networks (ANN)

Abstract

To depict hydrogeological variables and understand the physical processes taking place in a complex hydrogeological system, artificial neural networks (ANNs) are widely used as a good alternative approach to tedious numerical models. The aim of this study was to predict the dynamic fluctuations in piezometric levels in Nebhana aquifers (NE Tunisia) using ANNs. A correlation analysis carried out between piezometry, evapotranspiration and rainfall during the period 2000 to 2018 revealed that piezometric levels were influenced by monthly rainfall, evapotranspiration and initial water table level. These informative variables were used as input variables to train the ANN to predict future monthly water table levels for four hydrogeological systems. The minimal and maximal computed relative errors were 0.01 and 19.00%, respectively; root mean square error (RMSE) varied between 0.41 and 2.06; the determination coefficient (R2) ranged between 0.93 and 0.99; and the Nash–Sutcliffe (NASH) efficiency coefficient ranged from 85.32 to 97.82%. To test the generalization capacity of the developed ANN models, we used the ANNs to predict monthly piezometric levels for the period September 2016 to August 2018. The results were satisfactory for all piezometers. Indeed, the minimal and maximal computed RE were − 12.00 and 0.03%, respectively; RMSE was between 0.44 and 1.74; R2 varied between 0.95 and 0.98; the NASH coefficient ranged from 60.00 to 98.99%. These models developed in this study can be adopted for future groundwater level prediction to accurately estimate trends in piezometric levels as well as water pumping costs.