On May 6, the research team of DigitalTech Co., Ltd., in cooperation with the Department of Hydraulic Engineering of Tsinghua University, published a paper titled Perception-Enhanced Bidirectional Attention Networks for Short-term Time Series Forecasting of Water Level in Cascade Reservoir System in the Journal of Hydrology, a top international journal in hydrology and water resources. The paper proposed an innovative deep learning model named Perception-Enhanced Bidirectional Attention Network (PE-BANet), achieving major breakthroughs in prediction accuracy and computational efficiency for short-term water level forecasting of cascade reservoirs.
Cascade reservoir systems feature strong hydraulic coupling between upstream and downstream sections. Water discharge from upstream power stations reaches downstream areas with a dynamic time lag and is subject to the impacts of heterogeneous variables such as rainfall and dispatching orders. Traditional hydrodynamic models deliver sound physical interpretability yet carry heavy computational burden, making real-time parameter calibration difficult. Thus, the R&D team independently developed the PE-BANet model, adopting key technologies including Multi-Scale Attention Atrous Spatial Pyramid (MSAASP) and bidirectional attention mechanism on the basis of the ModernTCN framework, enabling the model to adaptively extract time series features across multiple time scales. The model can perceive short-term fluctuations and long-term water flow evolution rules like an expert, effectively addressing the prominent time lag during water flow propagation. Compared with mainstream models such as Transformer, MLP and RNN, PE-BANet takes a comprehensive lead in forecasting accuracy, reducing the average water level prediction error by 30%. Furthermore, it adopts an efficient pure convolutional architecture and delivers outstanding inference speed and computational efficiency, with its calculation time only one quarter that of mainstream counterparts.
Currently, the achievement has been launched on the Model Cloud, the cloud-based hydropower platform of DigitalTech, providing solid technical support for real-time dispatching of cascade hydropower stations across all river basins of CHN Energy. The lead researcher stated: "The algorithm integrating physical law perception and deep learning strengths can not only provide stable forecasting with a longer forecast horizon, but also ease the peak lag of traditional models amid abrupt water level changes, holding great significance for basin flood control safety and enhancing hydropower resource utilization efficiency."
