Enhancing Daily Scale Groundwater Level Prediction Accuracy by Integrating Physical Knowledge into LSTM Models

Abstract

Accurate groundwater level (GWL) prediction is crucial for water resource management. The nonlinear response of GWL to climate change and human activities makes daily scale prediction challenging. This study proposed three strategies to enhance daily scale GWL prediction accuracy by integrating physical knowledge into a long short-term memory (LSTM) model: incorporating hydrogeological parameters (LSTM-HP), considering precipitation recharge delay (LSTM-LAG), and embedding the relationship of GWL and precipitation (LSTM-MR-D). Results showed that LSTM-LAG achieved the most improvement, increasing the Nash–Sutcliffe efficiency coefficient (NSE) value by 0.01–0.42 during testing, followed by LSTM-MR-D and then LSTM-HP. However, combining strategies did not always enhance prediction accuracy. Approximately 50% of the observation wells that could have achieved accuracy improvements experienced a decrease instead. Additionally, hydrogeological parameters identified by LSTM can help calibrate a physics-based model, achieving satisfactory GWL prediction results with an NSE of 0.95. These findings offer practical solutions for improving GWL predictions and aiding groundwater management decision-making.