Study On Flood Forecasting Of Small And Medium-sized Rivers Based On Deep Learning And Hydrodynamic Model

Under the background of global climate change and the increase of extreme rainfall events,over standard floods occur frequently all over China.Due to the relatively weak flood control construction of small and medium-sized rivers in China,the losses caused by floods are very serious every year.The flood disaster control of small and medium-sized rivers has become a pain point and difficulty in flood control and disaster reduction in China.It is urgent to study the prediction method suitable for the flood characteristics of small and medium-sized rivers.In view of the demand for accuracy and timeliness of flood forecasting of small and medium-sized rivers,this paper studies the flood forecasting of small and medium-sized rivers based on the combination of deep learning and hydrodynamic model,constructs the watershed rainfall runoff forecasting model and river hydrodynamic model,and uses deep learning technology to improve the accuracy of rainfall runoff forecasting and hydrodynamic simuIation.The main results of the study are as follows:(1)The research on rainfall runoff prediction based on deep learning long-term and short-term memory network(LSTM)is carried out.As a data-driven model,LSTM has good nonlinear mapping ability,but it has the limitation of not considering the mechanistic factors of rainfall runoff process,which limits the improvement of its prediction accuracy.To solve this problem,this paper proposes an LSTM rainfall runoff prediction method considering the initial loss,combines LSTM with HEC-HMS semi distributed hydrological model.HEC-HMS is used to optimize the initial loss,and the initial loss is introduced into LSTM network to consider the impact of underlying surface closure on rainfall runoff.Taking the net rainfall data sequence obtained by deducting the initial loss of preliminary rainfall as the input and the runoff data sequence as the output,a rainfall runoff prediction model of LSTM basin considering the initial loss is constructed,which improves the accuracy of rainfall runoff prediction.(2)The simulation of river flood routing coupled with LSTM and HEC-RAS is carried out.Based on the hydrodynamic model(HEC-RAS),a river flood routing model coupled with deep learning model(LSTM)is established.The runoff process of LSTM rainfall runoff prediction is taken as the upstream flow boundary condition of hydrodynamic model,so as to realize the rapid prediction of the inundation range of downstream river from the upstream rainfall,and improve the timeliness of flood prediction.(3)Combined with deep learning convolution neural network(CNN),the roughness inversion of hydrodynamic model is carried out.Taking the section water depth data as the input and the roughness data of each river section as the output,a roughness inversion model based on CNN is constructed to eliminate the interference of subjective factors in manual roughness calibration and improve the simulation accuracy of hydrodynamic model.(4)Case application and analysis.Taking Yufu River in Jinan City,Shandong Province as an example,the application results show that,the prediction qualified rate of LSTM model without considering the initial loss is 60%,and the prediction qualified rate of LSTM model considering the initial loss is 80%.After introducing the initial loss,LSTM shows better prediction performance,the simulation results of HEC-RAS one-dimensional hydrodynamic model can basically reflect the change trend of real water surface profile,but the simulation error of water depth is high,After using CNN to retrieve roughness parameters,the error of water depth simulation is reduced,and the accuracy of hydrodynamic simulation is significantly improved,with the coupling of rainfall runoff prediction model and hydrodynamic model,the rapid prediction of flood inundation range of downstream river channel from rainfall data in the upper reaches of Yufu river is realized.Combining deep learning with flood forecasting to build a flood simulation and forecasting model suitable for small and medium-sized rivers is a potential development direction of flood forecasting for small and medium-sized rivers in the future.