Groundwater, as the key element of water resources, can play inevitably substantial role in managing groundwater aquafers. In fact, a ferocious demand for acquiring precise estimation of groundwater table is of remarkable significance for analyzing water resources systems. A wide range of artificial intelligence techniques were used to predict groundwater table with highly convincing level of precision. Hence, this investigation aims to present an integration of a neuro-fuzzy (NF) system and group method of data handling (GMDH) in order to forecast the ground water table (GWT). The NF-GMDH network has been improved by means of the particle swarm optimization (PSO) and gravitational search algorithm (GSA) as evolutionary algorithms. The proposed methods were developed using records of two wells in Illinois State, USA. For this purpose, datasets related to time series of GWT have been grouped into three sections: training, testing, and validation phases. Through training and testing phases, the efficiency of the NFGMDH methods were studied. The performances of proposed techniques were compared to the performance of radial basis functionneural network (RBF-NN). Evaluation of statistical results indicated which NF-GMDH-PSO network (R = 0.973 and RMSE = 0.545) is capable of providing higher level of precision rather than the NF-GMDH-GSA network (R = 0.969 and RMSE = 0.618) and RBF-NN (R = 0.814 and RMSE = 1.41). Also, conducting an external validation for the improved NF-GMDH models showed the most permissible level of precision.

Keywords