Graph-embedded time-interval-aware transformer for event-driven groundwater level forecasting

This paper proposes a groundwater level prediction method that integrates time-interval awareness with event-driven modeling, aiming to enhance model performance in non-stationary and abrupt hydrological processes. By incorporating event features into the attention mechanism, the framework effectively captures local mutations in groundwater level sequences, while probabilistic forecasting strengthens robustness against uncertain data. Experimental evaluations on eight monitoring wells from the California Department of Water Resources demonstrate that the proposed approach consistently outperforms multiple baseline models under diverse testing scenarios. Specifically, the method achieves an average reduction of 12.4% in MAE and 10.7% in RMSE, while the R2 metric exceeds 0.92. Even under conditions of high missing rates or perturbed timestamps, the model maintains stable predictive performance. These results confirm that the proposed framework delivers higher accuracy and stronger robustness in groundwater level forecasting under complex conditions, providing effective support for groundwater resource management and early warning applications.