De-noising groundwater level modeling using data decomposition techniques in combination with artificial intelligence (case study Aspas aquifer)

Abstract

Considering the recent significant drop in the groundwater level (GWL) in most of world regions, the importance of an accurate method to estimate GWL (in order to obtain a better insight into groundwater conditions) has been emphasized by researchers. In this study, artificial neural network (ANN) and support vector regression (SVR) models were initially employed to model the GWL of the Aspas aquifer. Secondly, in order to improve the accuracy of the models, two preprocessing tools, wavelet transform (WT) and complementary ensemble empirical mode decomposition (CEEMD), were combined with former methods which generated four hybrid models including W-ANN, W-SVR, CEEMD-ANN, and CEEMD-SVR. After these methods were implemented, models outcomes were obtained and analyzed. Finally, the results of each model were compared with the unit hydrograph of Aspas aquifer groundwater based on different statistical indexes to assess which modeling technique provides more accurate GWL estimation. The evaluation of the models results indicated that the ANN model outperformed the SVR model. Moreover, it was found that combining these two models with the preprocessing tools WT and CEEMD improved their performances. Coefficient of determination (R2) which indicates model accuracy was increased from 0.927 in the ANN model to 0.938 and 0.998 in the W-ANN and CEEMD-ANN models, respectively. It was also improved from 0.919 in the SVR model to 0.949 and 0.948 in the W-SVR and CEEMD-SVR models, respectively. According to these results, the hybrid CEEMD-ANN model is found to be the most accurate method to predict the GWL in aquifers, especially the Aspas aquifer.