Research On Flood Process Simulation And Forecast Based On Deep Learning

Flood disaster has always been one of the most serious natural disasters in our country.It threatens the safety of people’s lives and property,and hinders the sustainable development of society and economy.Accurate flood forecasting is an important non-engineering measure for flood control and disaster reduction.Especially in the past 40 years,the properties of the underlying surface in the middle reaches of the Yellow River have changed significantly,so the flood forecasting problem in the middle reaches of the Yellow River under changing conditions is very important.At the same time,with the advent of the era of big data,people can mine more in-depth information from a large number of hydrometeorological data,and the use of deep learning neural networks in the computer field to solve the problem of flood forecasting has become the focus of attention.Under the support of the National Natural Science Foundation of China "Research on the Key Technologies of Flood Forecasting in the Middle Reaches of the Yellow River Driven by Intensive Data(No.51979250)",this paper selects the Jingle River Basin in the middle reaches of the Yellow River as the research object,and constructs a neural network based on Long Short Term Memory(LSTM)and The deep learning flood forecasting model of temporal convolutional neural network(TCN)is combined with particle swarm optimization algorithm(PSO)to optimize deep learning hyperparameters,and the simulation results are compared with traditional machine learning models and physical models.The research contents and results are as follows:(1)Deep learning flood forecasting model modeling.From the input data to the model structure system,the modeling of the deep learning model and the process of extracting the rainfall and runoff features of the deep learning model are explained,and the LSTM and TCN flood forecasting models are constructed for simulation,and their simulation effects are compared.The study found that the simulation effects of the two deep learning models decreased with the increase of the forecast period.In the short forecast period,the two models showed better performance,with NSE above 0.98,and NSE above 0.7 in the long forecast period.TCN is higher than LSTM in overall simulation accuracy.(2)Construction of a deep learning flood forecasting model based on hyperparameter optimization.By establishing multiple sets of deep learning models with different hyperparameter combinations,the influences of various hyperparameters of LSTM and TCN on their flood simulation are obtained,and combined with particle swarm optimization,a deep learning flood forecasting model based on hyperparameter optimization is constructed.The hyperparameters are optimized within the set range,and finally a suitable combination of hyperparameters is found.In the long-term forecast period,PSO-LSTM and PSO-TCN are respectively 0.015 and 0.03 higher than those before optimization.After optimization,the simulated value is closer to the measured value,and the optimization effect is more obvious.(3)Contrast with traditional machine learning and physics models.Combined with the simulation results of deep learning model,traditional machine learning model(ANN)and physical model(EIESM)under different forecast periods,the simulation effects of each model in the flood process are compared.The results show that the simulation accuracy of all models decreases gradually with the increase of the forecast period.In a short forecast period,the deep learning model can achieve higher simulation accuracy than other models,the ANN model has a lower relative accuracy,and the EIESM model can reflect the flood process trend,but the accuracy is not as good as the neural network model.In the long-term forecast period,the deep learning model has a certain loss of accuracy,and PSO-TCN has a better simulation effect at this time.The ANN model has a serious decline in the simulation accuracy.EIESM has certain advantages over ANN models with lower accuracy.