Research On Short-term Precipitation Forecast Based On Spatiotemporal Series And Generative Networ

Precipitation nowcasting is a high-resolution forecast of rainfall and hydrometeor within 0-2hours in the future.Accurate precipitation nowcasting has always been an important and challenging global problem.Accurate precipitation nowcasting can provide information for the operations of many industries,including emergency services,energy management,flood warning systems,and air traffic control.In order to make precipitation nowcasting works effective in different application scenarios,nowcasting need to provide accurate predictions across multiple spatial and temporal scales.Traditional precipitation nowcasting mainly uses methods of numerical weather forecasting,statistical forecasting,and radar extrapolation.However,these methods all need to rely on a large amount of prior knowledge and artificial parameter settings,so it may cause problems such as low prediction accuracy and large limitations.In recent years,more and more scholars try to apply deep learning technology to the research of precipitation nowcasting with the rapid development of artificial intelligence technology.Therefore,this thesis carries out research on precipitation nowcasting based on deep learning,aiming to solve the shortcomings of existing methods and further improve the accuracy,clarity,and refinement of precipitation nowcasting.The main work and innovative points of this thesis are summarized as follows:(1)In response to the problem that existing methods cannot effectively capture long-range spatial dependencies in radar images and cannot model time information and spatial features uniformly,this thesis proposes a precipitation nowcasting model based on spatiotemporal memory cells and selfattention mechanism(SMSA-Conv LSTM)on the basis of Convolutional Long Short-Term Memory network(Conv LSTM).SMSA-Conv LSTM is able to model temporal information and spatial features in a unified manner,and further enhances its ability to extract long-range spatial dependencies.Experimental results prove that,the model can accurately predict the shape change and motion trajectory of radar echo images,and outperforms existing models in various evaluation metrics.(2)Aiming at problems such as insufficient clarity and unrealistic images predicted by the spatiotemporal sequence prediction model,this thesis proposes a generative precipitation nowcasting model that cooperates with Transformer GAN(Transformer Generative Adversarial Network)and SMSAConv LSTM.The model combines the image generation capability of generative adversarial networks and the spatio-temporal feature extraction capability of spatio-temporal sequence models.The evaluation results on the metrics for deep learning models and precipitation nowcasting show that while ensuring long-term prediction accuracy,the model can generate accurate and realistic images,and well meet the practical needs of precipitation nowcasting.Meanwhile,in response to the issue of unstable image quality generated by the generator,this thesis designed a new loss function for Transformer GAN and proved the effectiveness of the loss function through ablation experiments.In summary,the methods proposed in this thesis are innovative applications of deep learning in the field of precipitation nowcasting,providing new ideas and solutions for building accurate,clear,and refined precipitation nowcasting systems.These methods not only have strong generalization and broad application prospects,but also show excellent performance and effects in practical applications.Therefore,these methods are of great significance for promoting the development and application of precipitation nowcasting technology.