Optimization of gate operation for reservoir flood control based on deep learning ensemble forecasting

Research is needed for improving reservoir flood forecasting and operation to reduce flood-induced losses. Current reservoir flood control practices lack consideration of actual gate operations and dispatch scheme feasibility. This study proposes a refined flood operation optimization model, which takes into account inflow forecast uncertainties by utilizing ensemble forecasts derived from a combination of Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and Transformer deep learning models. Gate discharge states and their durations are taken as decision variables. A case study was conducted in the Jiaojiang River basin, Zhejiang Province. Results indicate that the ensemble forecasts are improved by 4.6% in accuracy compared to single model forecasts, and the forecast uncertainty is reliable for reservoir operation. The refined flood operation optimization model shows significant advantages over conventional methods, reducing the average maximum water level at the dam by 0.43 m and the average flood peak flow at downstream control section by 32.9 m³/s. The effectiveness of the optimized solution is less affected by inflow uncertainties. Overall, the proposed refined model is demonstrated to have the potential to provide valuable information to support reservoir flood control decision-making.