Research On Nowcasting Method Of Convective Initiation Based On Deep Learning

At present,people’s demand for nowcasting of convective initiation is increasing.With the continuous development of deep learning technology,it is possible to use deep learning methods to nowcasting of convective initiation.Although my country has made great progress in forecasting severe weather in recent years,traditional numerical weather forecasting methods still face greater challenges in nowcasting of convective initiation.In this thesis,we try to use deep learning methods to improve the timeliness and accuracy of the forecast model for the nowcasting of convective initiation.The work of this thesis is based on the horizontal projects controlled by the 14 th Research Institute of China Electronics Technology Group Corporation as follows:First,study the development and current situation for nowcasting of convective initiation.Investigate the current status of related application products and technologies,and at the same time summarize some of the existing problems at this stage,and explain the importance of using deep learning for nowcasting in the initial stage of convection activities.Second,multi-source data preprocessing.Through the analysis and processing of Sunflower 8 Meteorological Satellite data,East China radar data,reanalysis data,and primary tags of the stream,the generation of training data is completed and the standard database structure required for the experiment is established.Third,in view of the difficulty in modeling cumulus recognition in satellite cloud images,complex calculations,and weak generalization performance,a cumulus recognition method based on semantic segmentation and watershed algorithm fusion is proposed.Through neural network to learn the deep-level features of satellite cloud images,supplemented by watershed algorithm for correction,higher accuracy and stability can be obtained in the target cumulus cloud recognition task.Fourth,for the nowcasting model of convective initiation,the target cumulus correlation method based on modified Cost is used to correlate cumulus clouds,and the U-Net-based nowcasting model of convective initiation is proposed,and the data set in the established standard database is used to correlate cumulus clouds.The training of the forecast model and the introduction of the attention mechanism improve the accuracy and generalization of the nowcasting model,and reduce the false alarm rate of the nowcasting to a certain extent.Through the experimental verification of target cloud classification,it is found that through the fusion of semantic segmentation method and watershed algorithm,the main volume cloud target segmentation is correct,and it is less affected by the underlying surface,which improves the cloud classification effect to a certain extent.Secondly,the correlation experiment shows that the target cumulus correlation method based on modified Cost can effectively correlate cumulus targets.Finally,through the comparison of convective primary short-term forecasting experiments,U-Net has better generalization performance than long and short-term memory artificial neural networks.After introducing attention mechanism,random noise and correcting the loss function,the prediction model is improved.accuracy.Therefore,the nowcasting model proposed in this thesis has a good effect on the nowcasting of convection initiation in East China.