基于机理-深度学习耦合模型的流域洪水预报研究OA

Watershed flood forecasting is of great significance for flood control,emergency response,and water resources man-agement.To investigate the application of mechanism-deep learning coupled models in watershed flood forecasting,this study takes the upper reach of the Dongtiaoxi River Basin as the research object,construct the Xin'anjiang(XAJ)model,two deep learning models(LSTM and CNN-LSTM),and six mechanism-deep learning coupled models(PIML1/2/3-LSTM and PIML1/2/3-CNN-LSTM),and systematically evaluate the forecasting performance of these nine models for flood simulation.The results indicate that:① The XAJ model exhibits favorable overall stability,with an average Nash-Sutcliffe efficiency coeffi-cient(NSE)of 0.87 on the validation set.However,it shows systematic underestimation during high-flow periods,with absolute values of the relative error of total flood volume(RE)and the relative error of peak discharge(REP)both exceeding 15%.②Deep learning models(LSTM and CNN-LSTM)outperform the XAJ model during high-flow periods but yield slightly lower NSE values,decreasing by 5.7%and 10.3%compared with the XAJ model,respectively.③ Among the coupled models,the PIML1 type model that only incorporates the XAJ output discharge(QXAJ)achieves the best accuracy.Specifically,the RE and REP of the PIML1-LSTM model improve by 71.9%and 33.2%compared with the XAJ model.In contrast,the PIML2 type,which introduces soil moisture content(QSWV),shows a significant decline in performance,indicating that transferring certain in-termediate variables from physical models may inhibit the effectiveness of machine learning.④ The coupled models achieve signif-icant improvements in key flood characteristic indices at the cost of a slight reduction in overall fitting degree.It is recommended to prioritize the PIML1 architecture in practical applications,as it optimizes forecasting accuracy while maintaining physical inter-pretability.These findings provide a theoretical basis for optimizing the coupling of physical and data-driven hydrological mod-els.