基于双水源水量平衡约束的深度学习洪水预报模型OA

Global climate change,coupled with anthropogenic impacts,has led to frequent flood disasters,making high-accuracy and long-lead-time hydrological forecasting a critical support for enhancing disaster prevention and mitigation capabilities and ensuring water resource security.Existing artificial intelligence-based flood forecasting methods predominantly rely on extrapolation from single historical discharge series,failing to adequately consider the dynamic interaction mechanism of dual water sources from local precipitation and upstream inflow,as well as the water balance constraint.Consequently,these models exhibit weak representation of the physical mechanisms governing runoff generation and confluence,resulting in limited prediction accuracy during flash flood events.This paper proposed HydroFormer-TS,a deep learning flood forecasting model based on dual-source water balance constraints.By designing a network architecture that integrates dual-source information and embeds physics-guided features,the model was forced to satisfy the water balance principle,thereby enhancing the physical plausibility and interpretability of the simulation process.This approach advanced the modeling paradigm from purely"data-driven"to"data-physics fusion-driven".By taking the Beijiang River basin,a typical humid watershed in southern China,as the study area,multiple typical flood events and sudden extreme floods were selected for validation.The model demonstrates outstanding performance across all lead times,achieving a maximum Nash Sutcliffe efficiency coefficient of 0.94.It shows an improvement of 85.4%over the LSTM model and 38.2%over the iTransformer at the 72-hour lead time,while reducing the mean absolute percentage error to 23.05%for sudden flood events.This work provides a novel modeling perspective and technical pathway for high-precision hydrological forecasting.