Research On Flood Ensemble Prediction In The Upper Reaches Of The Huaihe River Based On Deep Learning

Flood disasters are highly destructive and sudden,accompanied by the characteristics of inaccurate prediction,which has endangered the safety of people’s lives and property and the stable development of society in many countries.Flood forecasting can provide basis and support for flood prevention work,so it is of great significance to build a flood prediction model with high forecast accuracy and strong applicability for flood prevention and disaster reduction.Most of the traditional hydrological models are modeled according to the physical mechanism,but the flood process is affected by many factors such as hydrology,vegetation,topography,underlying surface conditions under the river,etc.So,there are many model parameters,and the flood process itself is a complex nonlinear model,and a single mathematical and physical model cannot accurately describe this process.With the development of computer science and the advancement of hydrological informatization,various meteorological and hydrological data have become increasingly rich,and more scholars can model the data itself without being constrained by physical mechanisms,and modeling is easier,so data-driven flood forecasting models have emerged.Among them,deep neural networks are widely used.At present,most flood forecasting models based on deep learning use historical observation data as input,but meteorological forecast data can effectively improve the accuracy of flood ensemble forecasting.Therefore,based on historical hydrometeorological data and the ERA-Interim weather forecast data of European Centre for Medium-Range Weather Forecasts,this thesis constructs Convolutional Neural Networks(CNN),Gated Recurrent Unit(GRU),Temporal Convolutional Network(TCN)and Long Short-Term Memory(LSTM)models,conduct research on diurnal flood ensemble forecasting.In view of the fact that the above flood ensemble prediction based on deep learning is only based on site time series data,and does not fully consider the spatial heterogeneity of watershed characteristics,this thesis further uses GPM IMERG(GPM)global satellite precipitation products on the basis of adding precipitation forecasting,through convolutional long-term short-term memory network(Conv LSTM)and predictive recurrent neural network(Pred RNN)extracted the spatial characteristics of precipitation in the basin and constructed LSTM-Conv LSTM and LSTM-Pred RNN models to further improve the accuracy of flood ensemble forecasting.This thesis takes the basin above Xixian County in the upper reaches of Huaihe River as the research object,and studies the flood ensemble forecast for the two periods of1979～2016 and 2000～2016.The experimental results show that:(1)After adding precipitation forecast data to each deep neural network model(CNN,GRU,TCN,LSTM),Nash Efficiency Coefficients(NSE)are 0.803,0.861,0.878 and 0.901,respectively,which are 0.020,0.016,0.034 and 0.043 higher than the single forecast,and the root mean square error(RMSE)is 73.462,61.776,57.690 and 52.179,respectively,which is 3.574,0.474,7.387 and 10.126 lower than that of the single forecast.(2)Compared with the deep learning model used,the LSTM model has the best prediction effect,NSE reaches0.901,which is 0.098,0.040 and 0.023 higher than CNN,GRU and TCN,respectively;RMSE is 52.179,which is 21.283,9.597 and 5.511 lower than CNN,GRU and TCN,respectively.(3)The NSE result of LSTM model using time series data from 2000 to 2016 was 0.871,after considering the spatial characteristics of precipitation in the basin,the NSE of LSTM-Conv LSTM and LSTM-Pred RNN models reached 0.916 and 0.902,respectively,which were 0.043 and 0.031 higher than those of LSTM models,and RMSE were 50.167 and 54.446,respectively.The above experiments show that ensemble forecasting based on deep learning is better than single forecasting,and remote sensing precipitation products can effectively improve the prediction accuracy because they can better characterize the spatial heterogeneity of precipitation in the basin.This study can provide scientific basis for regional flood forecasting and water resources management.