Enhancing groundwater level prediction with a hybrid deep learning model in Jinan City, China

Accurate prediction of groundwater levels (GWL) is critical for sustainable utilization and scientific management of groundwater resources. However, precise forecasting of GWL fluctuations faces significant challenges due to the complex nonlinear coupling effects of hydrogeological conditions and hydro-meteorological factors. In recent years, research on GWL prediction based on deep learning models has become a cutting-edge topic in the field of hydrogeology. This study focused on Jinan City, China, and constructed a novel hybrid deep learning model that integrates graph neural networks to capture spatial relationships and recurrent neural networks to model temporal dynamics, effectively learning the complex spatio-temporal patterns in the data, namely the Spatio-Temporal Graph Prediction Model (STGPM). Our approach uniquely captures both hydrological connectivity between monitoring wells and multi-scale temporal dependencies, overcoming key limitations of conventional time-series models. Comparative experiments demonstrate that STGPM outperforms the benchmark models on the test set, achieving the lowest prediction errors (MAE = 0.039, RMSE = 0.052) and the highest coefficient of determination (R²=0.988). Notably, for the monitoring well data not involved in model training, the STGPM still maintains excellent predictive accuracy (MAE = 0.062, RMSE = 0.087, R²=0.980), demonstrating the model's strong generalization ability to unmonitored locations. This study provides water resource managers with a reliable decision-support tool for sustainable groundwater management and spring conservation strategies. The proposed methodological framework also offers a transferable solution for addressing various environmental forecasting challenges characterized by spatial heterogeneity.