A study on the GRU-LSTM hybrid model with dual attention mechanism for reservoir water level forecasting

Abstract Reservoir water level forecasting is essential for flood control and water resource optimization, yet remains challenging due to the nonlinear, non-stationary, and multi-scale nature of hydrological processes. This paper proposes a hybrid deep learning model (GLA-GRU-LSTM) that integrates GRU and LSTM with dual attention mechanisms to capture both long-term trends and short-term fluctuations in water level dynamics. The global attention module extracts long-range dependencies across the entire sequence, while the local attention module focuses on abrupt changes within a recent temporal window. The model was evaluated using ten-year (2014–2023) hydrometeorological data from Yefan Reservoir, China, and compared against SVR, LSTM, CNN-LSTM, Transformer, TFT, and RFM_Trans. For 7 day forecasting, GLA-GRU-LSTM achieved NSE of 0.934 and RMSE of 1.647, outperforming Transformer (NSE = 0.921) and TFT (NSE = 0.916). For 30 day forecasting, it maintained superior performance (NSE = 0.807, RMSE = 2.773), with the performance gap widening as forecast horizon extended. Ablation studies confirm that performance gains stem from synergistic integration of GRU-LSTM hybridization and dual attention, not merely increased parameterization. Statistical significance tests (Diebold-Mariano, p < 0.05) and bootstrap confidence intervals confirm the robustness of improvements. The proposed framework provides an accurate and interpretable approach for multi-scale reservoir water level forecasting, with potential applications in operational water management.