Reservoir group runoff forecasting based on machine learning and conceptual reservoir operation module based on storage zones

Accurate runoff forecasting in river basins regulated by cascade reservoirs is essential for water resources management and flood risk reduction. However, conventional approaches struggle to simultaneously capture upstream cascade effects and complex reservoir operation behaviors. This study focuses on the Eping-Longbeiwan-Pankou reservoir system in the Duhe River Basin and develops a three-stage integrated runoff forecasting framework that couples an Attention-LSTM model with a Conceptual reservoir operation module based on Storage Zones (CSZ). The framework first predicts inflows to upstream reservoirs, then employs the CSZ module to simulate operation processes and transform inflows into regulated outflows, which subsequently serve as boundary conditions for downstream runoff prediction. Daily data from 2021 to 2024 were partitioned chronologically into training, validation, and testing sets. Results demonstrate that the proposed framework achieves NSE values of 0.85-0.94 for 1- to 7-day lead times, with RMSE controlled within 67-134 m³/s and relative errors below 2.68%, substantially outperforming benchmark models such as standard LSTM. Ablation experiments quantitatively reveal that cascade information contributes 44.4% to performance gains, the attention mechanism contributes 29.6%, and the CSZ module contributes 14.8%. Cascade effect analysis further indicates that upstream reservoirs collectively account for 64.9% of improvements in downstream prediction, with Eping contributing 27.1%, Longbeiwan 13.8%, and synergistic effects 7.5%. For representative flood events, the model attains an NSE of 0.97 for major floods (Event A) with peak flow relative errors below 6%, and an NSE of 0.68 for moderate floods (Event B). Variable importance analysis identifies historical flow and upstream outflow as the most critical predictors, while precipitation variables exhibit maximum contribution at 1-2 day lags. By integrating predictive accuracy, physical consistency, and interpretability, the proposed framework provides robust scientific support for intelligent multi-reservoir operation and flood early warning systems.