GROUNDWATER LEVEL PREDICTION USING DEEP RECURRENT NEURAL NETWORKS AND UNCERTAINTY ASSESSMENT

Abstract

Groundwater is one of the most important sources of regional water supply for humans. In recent years, several factors have contributed to a significant decline in groundwater levels (GWL) in certain regions. As a result of climate change, such as temperature increase, rainfall decrease, and changes in relative humidity, it is necessary to investigate and model the effects of these factors on GWL. Although a number of researches have been conducted on GWL modeling with machine learning (ML) and deep learning (DL) algorithms, only a limited number of studies have reported model uncertainty. In this paper, GWL modeling of some piezometric wells has been conducted by considering the effects of the meteorological parameters with Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) algorithms. The models were trained on one piezometric well data and predictions were executed on six other wells. To perform an uncertainty assessment, the models were run 10 times and their means were calculated. Subsequently, their standard deviations were considered to evaluate the outcomes. In addition, the prediction power of the models was validated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Normalized Root Mean Square Error (NRMSE), and R-Squared (R2). Finally, for all the six wells that did not participate in the training phase, the prediction functions of the trained models were run 10 times and their accuracy was assessed. The results indicate that LSTM (R2=95.6895, RMSE=0.4744 m, NRMSE=0.0558, MAE=0.3383 m) had a better performance compared to that of GRU (R2=95.2433, RMSE=0.4984 m, NRMSE=0.0586, MAE=0.3658 m) on the GWL modeling.