Research On The Method Of Nowcasting Of Severe Convective Weather Precipitation Based On Deep Learning

Strong convective weather has the characteristics of rapid evolution,short duration,and vigorous destructive power,and often poses a serious threat to people's lives and property.Therefore,it is significant to study the nowcasting methods of stong convective weather.In meteorological operations,Doppler weather radar is usually used to collect the weather data and estimates the atmospheric wind field and precipitation of the severe convective weather system.In this thesis,deep learning-based methods are proposed to forecast the precipitation within half an hour through the Doppler weather radar data.In order to process and build the Doppler weather radar data.The RGB value in the radar image is converted to the corresponding dBZ value according to the radar echo map key.Then the converted image is cropped to obtain the active echo area.After sorting and grouping all the preprocessed images,a radar echo dataset is obtained.Based on the experimental datasets,two forecast methods are proposed.Firstly,a novel nowcasting prediction algorithm is proposed by applying the attention mechanism to current deep learning prediction methods.The key of this method is to learn the feature map obtained by the convolution layer.The scores and weights of the feature maps are computed to obtain the critical feature information,and extract features based on the ConvLSTM module.Experimental analysis on a specific weather case domenstrates that the proposed method can improve the performance of echo intensity prediction.Secondly,the nowcasting prediction algorithm is improved by exploiting the convolutional Auto-Encode structure.The attention mechanism and convolutional AutoEncode are fused to improve the prediction ability of the model.The key of the method is to learn the essential feature maps extracted by the encoder combined with the attention module.The sequential features between feature maps are extracted,and the features are reconstructed reversely by the decoder.This method can improve the learning ability of the model regarding to the echo's moving tracks and changes.Experiments on the test data show that our method can reduce the missing reports rate and significantly improve the critical success index.On the basis of the above research,strong convective weather forecast system has been realized.Through analyzing the user demands based on the time sequence diagram,system functions are determined.The system consists of four modules: user management,nowcasting prediction,prediction record and score calculation.The system is developed by the Django framework and can directly call the Kears framework.Morevore,message queue is adopted to deal with prediction tasks,which improves the system performance and user experience.