Integrated probabilistic forecasting framework for long-term reservoir outflow through dynamic coupling of meteorological–hydrological–engineering processes

Abstract

To address challenges of uncertainty and complex multi processes coupled modeling in long-term streamflow forecasts after reservoir operation, this study proposed an integrated data–mechanism-driven framework for long-term probabilistic reservoir outflow forecasting. First, the Dempster–Shafer evidence theory identifies key predictors. Then, a combined quantile regression–convolutional neural network–bidirectional long short-term memory (QRCNN-BiLSTM) model is developed for probabilistic inflow forecasting. Thereafter, multi-quantile scenarios generated by inflow forecasts integrated with operation-related factors are used as input, and physical mechanisms are integrated in the loss functions. A LightGBM model containing reservoir operation knowledge forecasts probabilistic outflow, thereby realizing simulation prediction of natural runoff to regulated runoff. Applied to the Bengbu Sluice on the Huai River, the results were as follows. (1) The QRCNN-BiLSTM model reduced the root mean square error (RMSE) by 6.8 % and improved the Nash–Sutcliffe efficiency by 6.0 %, outperforming the CNN-BiLSTM benchmark model in the prediction of inflow. It showed a higher coverage rate (CR) and narrower average relative bandwidth (RB) compared to the QR neural network (QRNN) benchmark model, with a 24.8 % reduction in mistaken deviation (MD). (2) The LightGBM model outperformed the LSTM benchmark model, reducing the RMSE, continuous ranked probability score, and MD by 3.5 %, 6.0 %, and 7.5 %, respectively, while also achieving better CR and RB values in the prediction of outflow. (3) The integrated QRCNN-BiLSTM–LightGBM model outperformed the QRNN–LSTM model within 1–2-month lead time. The proposed framework offers a more accurate, reliable, and robust probabilistic forecasting solution for water resource optimization.