Research On Runoff Prediction In The Upper Huai River Basin Based On Deep Learning

Under the background of global warming,many studies have confirmed that the global water cycle is gradually strengthening.Coupled with the influence of a large number of reservoirs and other human activities,watershed runoff process is undergoing complex and far-reaching changes,and the runoff prediction is also becoming more difficult.At the same time,extensive research and application of deep learning techniques in various fields have also promoted the intersection and integration of deep learning with hydrology.Using deep learning to solve difficult problems in runoff prediction has become a current research hotspot.The upper of Huai River basin is located in the middle of China,spanning Henan and Anhui,two provinces with large grain and large population.The special geographical location and the important influence on national food security make the prediction and management of runoff in the upper of Huai River basin more important.Therefore,this paper selects the upper of Huai River basin as the research object,and conducts research on the runoff prediction of the basin based on deep learning in this region in order to facilitate the management of flood control and disaster reduction and water resources in the basin,and explore the development direction of cutting-edge technology in runoff prediction,providing important technical support and reference for the hydrological work of the basin.The main research content of this paper is as follows:(1)The linear regression method,Mann-Kendall test,and significance test are used to analyze the trends of temperature,evapotranspiration,precipitation,and runoff depth in the upper Huai River basin from 1951 to 2016 on different time scales,while remote sensing images are used to analyze the changes in land use in the upper of Huai River.The Mann-Kendall test results show that the meteorological and hydrological characteristics of the upper of Huai River basin have not changed dramatically in 65 years,in which the precipitation and runoff depth showed a weak downward trend,but the temperature increased significantly and the evapotranspiration decreased significantly.The decrease of runoff depth may be related to the increase of forest water storage capacity caused by the obvious expansion of forest area.(2)In order to improve the prediction results of both overall runoff and extreme runoff,this paper separately reorganizes the features for meteorological stations in the basin,constructs an Enhanced-LSTM(ELSTM)model based on Long Short-Term Memory(LSTM),and constructs loss functions PET and PES to improve the prediction results of extreme runoff.The results showed that: The ELSTM using PET loss function has the best performance in the overall runoff prediction,with NSE up to 0.924,while the ELSTM using PES loss function has the best result in the extreme runoff prediction,and the qualified rate of flood prediction is 92.3%.The NSE of data-driven models including SVR,GRU and ANN and lumped hydrological models such as AWBM and Sacramento is only up to 0.904,and the qualified rate of flood forecast does not exceed 70%,which shows the superiority of ELSTM in the prediction of overall runoff and extreme runoff.(3)To explore the impact of reservoirs on runoff prediction,this study first directly incorporated reservoir data as features into models such as ELSTM,SVR,and GRU using traditional methods.It was found that the improvement in runoff prediction with the inclusion of reservoir data was limited and it reduced the accuracy of flood forecasting.In order to naturally incorporate reservoir information into runoff prediction models,this study treated hydrological stations and reservoirs as nodes and constructed a directed graph neural network based on the watershed routing topology.The DAGNN(Directed Acyclic Graph Neural Network)model was constructed using the architecture of an Variational Auto-Encoder to perform synchronized runoff prediction for multiple stations.The results showed a clear upward trend in hydrological station runoff predictions from upstream to downstream.The runoff prediction results for Xixian hydrological station,located downstream where more information is converged,far exceeded those for Dapoling and Changtaiguan in the upstream,with NSE up to 0.928,and the qualified rate of flood prediction could also reach 80%.Model comparison results indicated that although the DAGNN model,which incorporates reservoir information,was not as good as ELSTM,performed better than other models(SVR,GRU,ANN),particularly excelled in predicting extreme runoff.