Research On Runoff Prediction And Interpretability In The Yangtze River Basin Based On Deep Learning Method

Watershed streamflow prediction plays a vital role in water resources management,flood control,energy development,drought resistance.Data-driven rainfall streamflow model predicts streamflow by mining the relationship between driving factors and target values in historical data,and has high flexibility in application.On the one hand,the current deep learning is mainly used in the field of hydrology-runoff prediction as a time series modeling algorithm,but the simple time series model structure often limits the description of the spatial changes of hydrometeorological elements.Whether the deep learning model can improve the runoff prediction effect of watersheds remains to be studied.On the other hand,the hydrological significance of deep learning streamflow model remains unclear.Therefore,the research,which takes the Yangtze River sub-basin as the research object,include quantifying the influence of precipitation uncertainty on the data-driven modeling using multiple sets of precipitation,improving the accuracy of streamflow prediction using the LSTM variant model system,and revealing the hydrological significance of deep learning streamflow model in both temporal and spatial dimensions.Focusing on the above objectives,the main conclusions as follows:(1)Based on multiple sets of precipitation products,the runoff in the study areas is predicted by the LSTM models.The contribution of precipitation before the peak flow to runoff simulation is quantified which explores the hydrological interpretation of LSTM runoff model in the time dimension,and analyzes the impact of multi-source precipitation data on runoff prediction is analyzed.The results show that the NSE values of the model predictions driven by CMA and CN05 precipitation data in each basin are about 0.8,and NSE driven by CMFD,ERA5,and GLDAS reanalysis precipitation data are unstable,around 0.6-0.8.By using the integral gradient method,it is found that in the high-altitude areas of the upper reaches of the Yangtze River,the combined effect of snowmelt runoff,historical precipitation,and recent precipitation is the main cause of floods in the Jinsha River and the Minjiang River,among which the proportion of snowmelt runoff control in the Jinsha River flood events is46.81%.In the Wujiang River and Xiangjiang River,the combined effect of historical and recent precipitation is the main cause of inducing basin floods,while in the Ganjiang River,the flood events is controlled by historical precipitation events.The LSTM model can improve the runoff prediction effect by simultaneously adapting to multiple sets of precipitation data.(2)By constructing of a CNN-LSTM model which considered spatial characteristics of hydrometeorological elements,the hydrological interpretation of the CNN-LSTM model in the spatial dimension is explored.The study uses visualization of convolutional layers to analyze the relationship between model weight and precipitation feature distribution.The results show that the predictive performance of the CNN-LSTM model,driven by various precipitation datasets,is generally better than that of the LSTM model in various watersheds,confirming that utilizing spatial information of precipitation is conducive to improving the predictive performance of deep learning models.Through visualization of the model’s convolutional layers indicated that there is a high similarity between the weight distribution of the convolutional layers and the spatial distribution map of multi-year average precipitation in the Jinsha River and Minjiang River.In the Xiangjiang River and Ganjiang River,there is a high correlation between the weight distribution of the model’s convolutional layers and the spatial distribution of multi-year average monthly maximum precipitation in the basin.Precipitation in areas with high convolutional layer weights has a greater impact on the results.These results provide a new approach to improving the predictive ability of the model for runoff.