Groundwater level estimation using improved deep learning and soft computing methods

Abstract

Estimating groundwater level (GWL) is an important issue for planning and managing available water resources. This study uses monthly data from 86 observation wells from Mashhad Plain in Iran. A principled hierarchy method was used for the first time. In this regard, the K-means-GA method was used for clustering the considered wells. In each cluster, Principal Component Analysis (PCA) was employed to remove extra-loading observation wells. The presented study examines the accuracy of a new deep learning method, Long Short-Term Memory (LSTM), with Grey Wolf Optimization (GWO) (LSTM-GWO hybrid model) in modeling the GWL. The outcomes of the LSTM-GWO are compared with the enhanced artificial neural network (ANN), hybridized with GWO (ANN-GWO), and standalone ANN in the estimation of GWL. The results revealed that the LSTM-GWO method has a better ability to estimate GWL than the ANN-GWO and ANN methods. In the testing phase, by using the GWO the mean absolute average (MAE) of the ANN-GWO models decreased by at least 30% compared to the standalone ANN models. In addition, for ANN-GWO models the CA parameter which combines the root mean squared error (RMSE), MAE, and R² decreased by at least 15% in the testing phase compared to the standalone ANN model. The ANN is the least accurate method to estimate monthly GWL. Hybrid model LSTM-GWO almost 23% improved the GWL estimations compared to previous research in terms of coefficient of determination, R².