Research On The Extrapolation Of Meteorological Radar Echo Based On Recurrent Network

Nowcasting refers to detailed monitoring of the current weather development,mainly forecasting the weather conditions within 2 hours.Nowcasting focuses on the monitoring of sudden natural disasters such as hail,squall line,typhoon,heavy precipitation and thunderstorm,which is of great significance to the safety of people’s lives and property.At present,nowcasting mainly uses historical radar echo images to effectively extrapolate future radar echo images.Therefore,radar echo extrapolation algorithms have become the key to nowcasting.Since the1950 s,people have studied radar echo extrapolation and proposed a series of traditional methods.the TREC and the optical flow method have been put into specific business applications.However,traditional algorithms based on the Lagrangian conservation,such as the optical flow method,when encountering local strong convective weather,the echo changes rapidly and cannot meet the conservation conditions,resulting in a poor result.Moreover,traditional methods cannot make use of the existing massive meteorological data.As deep learning has been effectively applied in different fields,people have begun to study and propose a series of radar echo extrapolation models based on deep learning algorithms.Although the extrapolation effect has been improved,there are still extrapolated images that are blurred and distorted.In response to this problem,this disseration has mainly done the following work:First of all,this disseration proposes a radar echo extrapolation model based on gradient prediction,Ghu LSTM,based on the existing deep loop network.By studying the radar echo extrapolation model,it is found that the problem occurs in two points: the cyclic neural network unit’s ability to model high-order non-stationary features of the Spatio-temporal information is insufficient;the gradient disappears during training.Therefore,this paper adds two new structures to the model: In the extracted high-level feature information,the Spatio-temporal memory unit variant is introduced,and the differential signal between adjacent recursive states is used to process the high-order non-evolution of the radar echo.In order to alleviate the trend of rapid disappearance of gradients in the extrapolation,a gradient highway structure is added to the network.Through comparative experiments,it is proved that these two structures can effectively improve the accuracy of extrapolation.Secondly,through the analysis of the radar echo image generated by the model Ghu LSTM,the result still has the problem of insufficient clarity and lack of detail.Therefore,this dissertation proposes a deep radar echo extrapolation model based on cascade structure,Co RNN.The model introduces a cascade structure on the basis of Ghu LSTM,and enhances the depth of the model,which can effectively obtain the spatial correlation characteristics and shortterm dynamic characteristics of the radar image sequence.Through comparative experiments,this dissertation proves that the cascade structure can improve the quality of the generated image.