基于分布式产流要素和时空深度学习算法的径流后处理研究OA北大核心CSTPCD

Accurate simulation of streamflow is a crucial prerequisite for water resources management and regional integrated policy making.In order to improve the accuracy of streamflow simulation,this study takes Yonganxi Riv-er Basin in Taizhou,Zhejiang Province as the study area.A CNN-LSTM spatio-temporal post-processing model by coupling CNN with LSTM is proposed based on the measured daily discharge data at Baizhi'ao Station from 2010 to 2019 and hydrological fluxes simulated by the Grid-HBV model.We construct two post-processing models,namely CNN-LSTM with single flux(s-CNN-LSTM)and CNN-LSTM with double fluxes(bi-CNN-LSTM).Their performance is compared and analyzed with a benchmark model(s-LSTM).The results show that the NSE of the Grid-HBV model during the calibration and validation periods are 0.78 and 0.81,respectively,indicating an over-all good runoff simulation.However,there are underestimation in medium and high flow and overestimation in low flow simulations.After post-processing,the NSE of s-LSTM in the two study periods are 0.87 and 0.85,with an increase of 11.2％and 5.8％,and the NSE of s-CNN-LSTM are 0.90 and 0.89,with an increase of 14.6％and 10.9％.The NSE of bi-CNN-LSTM in the two study periods both reach 0.92,with an increase of 17.2％and 14.2％.Compared to the s-LSTM model,the bi-CNN-LSTM model presents a further enhancement of 6.0％and 8.4％in accuracy.In addition,the bi-CNN-LSTM model can markedly improve the defects of original simulation in the high,medium and low flows.For four typical flood events,the bi-CNN-LSTM model has the best post-processing effect,which reduces the flood peak error by 36.6％on average,the s-LSTM model and the s-CNN-LSTM model reduces the flood peak error by 19.3％and 30.3％on average.In summary,the CNN-LSTM model based on dis-tributed hydrological fluxes has a good ability of streamflow post-processing,which can significantly improve the streamflow simulations of hydrological models.