GR4J与深度学习融合的径流模拟方法研究OA

To address the limited adaptability of traditional conceptual hydrological models in simulating complex runoff processes,this study proposes a hybrid modeling framework,termed GR4J-DL,which integrates the GR4J model with deep learning architectures.The proposed framework retains the runoff generation structure of GR4J while employing three neural network architectures—LSTM,TCN,and TKAN—to represent nonlinear flow routing processes,thereby enhancing the model's ability to capture temporal dynamics and spatial structural information.The Enshi hydrological station in the upper Qingjiang River basin is selected as a representative case,and daily meteorological and runoff data from 2010 to 2018 are used for model training and validation to evaluate the performance of different hybrid configurations.The results indicate that,except for the GR4J-LSTM configuration,the GR4J-DL hybrid models achieve consistent improvements in overall simulation accuracy.Among them,the TKAN-based model performs best,with an NSE of 0.757 3,RMSE of 43.67 m3/s,KGE of 0.800 and TPE of 0.212,demonstrating a stronger capability in representing nonlinear runoff behavior.Further feature contribution analysis based on the SHAP method reveals that the intermediate state variables generated by the GR4J runoff generation module primarily act as physical constraints and auxiliary information,which contributes to improving model accuracy.Overall,this study broadens the applicability of deep learning in runoff modeling and provides a promising pathway for enhancing the simulation of complex runoff processes.