The research team of CHN Energy Digital Tech Co,. Ltd., together with the Department of Hydraulic Engineering at Tsinghua University, published a paper on May 6 in the internationally renowned magazine Journal of Hydrology, introducing the Perception-Enhanced Bidirectional Attention Network (PE-BANet) deep learning model. The model achieves major breakthroughs in both forecasting accuracy and computational efficiency for short-term water level prediction in cascade reservoirs.
Main findings of the research
Affected by hydraulic interactions between upstream and downstream reservoirs as well as multiple external factors, cascade reservoirs pose significant challenges for traditional models. Although conventional approaches feature clear physical logic, they require heavy computational workloads and are difficult to calibrate in real time, making them unable to meet dispatching demands under complex operating conditions. Verified through practical application, the newly developed PE-BANet model optimizes algorithm design to reduce average water level prediction errors by 30 percent compared with mainstream models, while cutting computation time to one quarter of that required by comparable systems. The model effectively addresses forecasting difficulties caused by delayed water flow propagation, achieving industry-leading performance in both prediction accuracy and operational efficiency.
The achievement has now been integrated into the company’s cloud-based hydropower platform, providing reliable technical support for the real-time dispatching of cascade hydropower stations across the Group’s river basins. The model delivers stable forecasts with longer lead times and alleviates the peak-lag phenomenon during sudden water level fluctuations. It is expected to play an important role in safeguarding flood control safety and improving hydropower resource utilization efficiency, while also offering a valuable reference for technological innovation in hydrological time-series forecasting.
