A comparison of groundwater level forecast modelling using deep learning models combined with EMD and EEMD in Cangzhou, China

Abstract

Accurate groundwater level forecasting is crucial for sustainable water resource management and environmental protection. The main objective of this study was to develop hybrid models combining deep learning models with data processing techniques for groundwater level forecasting. Two groundwater level observations in Cangzhou City, Hebei Province, China, were used as case studies. First, the empirical mode decomposition (EMD) and ensemble empirical mode decomposition (EEMD) were used to pre-process the dataset. Then, two dynamic deep learning techniques, namely long short-term memory (LSTM) and bidirectional LSTM (BiLSTM), were used as the base model. The hybrid models combined two deep learning algorithms (LSTM and BiLSTM) with EMD and EEMD, respectively, yielding 4 models: EMD-LSTM, EEMD-LSTM, EMD-BiLSTM, and EEMD-BiLSTM. Five evaluation metrics, including the root mean squared error (RMSE), the mean absolute error (MAE), the Nash-Sutcliffe efficiency (NSE), the Kling-Gupta efficiency (KGE), and the coefficient of determination (R2), were used to evaluate the proposed models. All the proposed models could achieve satisfactory results, with R2, NSE, and KGE values greater than 0.87. The hybrid models performed better than LSTM and BiLSTM, with reductions of up to 75% in RMSE and 71% in MAE. Furthermore, the EEMD-LSTM and EEMD-BiLST models achieved relatively lower RMSE and MAE values and higher NSE and KGE values for both wells, compared with the EMD-LSTM and EMD-BiLSTM models. The two decomposition algorithms significantly improved the performance of the LSTM and BiLSTM models, especially the EEMD. This study demonstrated that the EMD and EEMD models, combined with LSTM and BiLSTM models, can be a powerful tool for addressing non-stationary, complex groundwater level time series data.