Groundwater level prediction with meteorologically sensitive Gated Recurrent Unit (GRU) neural networks

Abstract

Precise estimation of groundwater level (GWL) fluctuations has a substantial effect on water resources management. In the present study, to forecast the regional mean monthly time series groundwater level (GWL) with a range of 4.82 (m) in Urmia plain, three different layer structures of Gated Recurrent Unit (GRU) deep learning-based neural network models via the module of sequence-to-sequence regression are designed. In this sense, 180-time series datasets of regional mean monthly meteorological, hydrological, and observed water table depths of 42 different monitoring piezometers during the period of Oct 2002–Sep 2017 are employed as the input variables. By using Shannon entropy method, the most influential parameters on GWL are determined as regional mean monthly air temperature (Tam), precipitation (Pm), total (sum) water diversion discharge (Wdm) of four main rivers. Nevertheless, Cosine amplitude sensitivity analysis confirmed Tam as a dominant factor. For preventing overfitting problem, an algorithm tuning technique via different kinds of hyperparameters is operated. In this respect, several scenarios are implemented and the optimal hyperparameters are accomplished via the trial-and-error process. As stated by the performance evaluation metrics, Model Grading process, and Total Learnable Parameters (TLP) value, the innovative and unique suggested model (3), entitled GRU2+, (Double-GRU model coupled with Addition layer (+)) with seven layers is carefully chosen as the best model. The unique suggested model (3) in the optimal hyperparameters, resulted in an R2 of 0.91, a total grade (TG) of 7.76, an RMSE of 0.094 (m), and a running time of 47 (s). Thus, the model (3) can be certainly employed as an effective model to forecast GWL in different agricultural areas.