Study On Doppler Radar Echo Extrapolation Algorithm Based On Deep Learning

Strong convective weather is rapidly evolving and destructive,and is one of the most devastating weather conditions that seriously threatens people's lives and property.Weather nowcast with radar echo extrapolation and numerical forecasting is the main means of monitoring and early warning of strong convective weather.Radar echo extrapolation is to predict future radar echo data based on the current radar echo data,it is the main method of weather nowcast within 2 hours.Traditional radar echo extrapolation method performs well on stable echo,and can track the change of echo motion accurately.However,the prediction accuracy is greatly reduced when radar echo changes rapidly.In addition,the prediction time of traditional radar echo extrapolation method is short,the accuracy decreases rapidly with the increase of predition time.With the gradual deployment of the new generation of Doppler weather radar in China,how to improve radar echo extrapolation algorithms and use the Doppler weather radar echo data effectively to get better accuracy and longer prediction time is a significant research direction.This paper first analyzes the working mode and echo data of the new generation Doppler weather radar CINRAD-SA,introduces and analyzes the widely used radar echo extrapolation algorithms.According to the sequential characteristics of radar echo extrapolation,this paper introduces the RNN network which is used to solve sequential problems in Deep Learning.After studying and analyzing the principle and characteristics of the RNN network,the LSTM model of the RNN network is applied to the radar Echo extrapolation.Finally,according to the applying of LSTM model,this paper proposed the RET-RNN model based on the LSTM model,and optimized the structure and the superparameters of the RETRNN model.In meteorology,the reflectivity factor is associated with the intensity of precipitation,and the reflectivity factor map generated by radar echo is an important tool for measuring short-term temporary precipitation.In this paper,we use the reflectance factor graph as extrapolated target,train RETRNN model with historical echo data,then extrapolate and evaluate the trained RET-RNN model.The results show that the echo extrapolation algorithm based on the RET-RNN model has a longer prediction time than traditional radar echo extrapolation algorithms.