Research On Typhoon Movement And Intensity Forecasting Based On Deep Learning

China is located in the Western North Pacific basin where is the most active typhoon region in the world.Every year during the summer and autumn typhoon seasons,China 's coastal areas will be suffered varying degrees of economic loss and casualties due to typhoons attack.Predicting the path and intensity of typhoons accurately is very important for the prevention and control of meteorological disasters in China 's coastal areas.At present,statistical typhoon prediction based on deep learning methods is still in the exploration stage.Most of the current research is almost based on the 2D characteristics of the typhoon and does not consider well of the 3D characteristics of the typhoon.Additionally,the fusion of information of 2D and 3D typhoon model is lacking,which is likely to cause lo w prediction accuracy to both tasks.This article focuses on the fusion and mining of multi-modal heterogeneous time-series features in typhoon forecasting and studies the methods of typhoon path prediction and typhoon intensity prediction based on deep l earning.The main research works and results include the following three aspects.In terms of typhoon path prediction,for the quality evaluation of the typhoon dataset in the preprocessing stage,I propose a method for evaluating the CLIPER characteristics of the dataset using Autocorrelation Coefficients.For the problem of selecting the isobaric planes of the typhoon path during the model training stage,I introduce Residual Channel Attention(Re CA)to automatically select high-response isobaric planes.In the model training phase,to solve the problem that traditional Conv LSTM cannot extract large-scale deep nonlinear features,I propose an iterative convolution GRU method(Multi-Conv GRU).Finally,I integrate the twodimensional and three-dimensional chronological heterogeneous features of typhoon under the wide&deep framework.Experiments show that the method proposed in this paper can effectively integrate the multi-modal typhoon timing features,which greatly improves the 24-hour typhoon path prediction accuracy.In terms of typhoon intensity prediction,for the problem of typhoon intensity component learning at the model training stage,I propose a Spatial Attention Based Fusing Network(SAF-Net)model based on spatial attention mechanism.The model decomposes the typhoon intensity into U Wind variable and V Wind variable,and introduces a spatial attention mechanism to each isostatic surface to extract the high-response region of the wind speed component,and then fuses the two wind speed components through convolution fusion.Achieve the goal of predicting the closing speed by learning each wind speed component.Experiments show that using the SAF-Net model proposed in this paper can improve the accuracy of 24-hour typhoon intensity prediction.Also,we designed and implemented a typhoon path and intensity visualization system based on the Flask framework.The front-end responsive layout was implemented through the Bootstrap framework,and the front-end page linkage and back-end updates were completed using asynchronous communication mechanisms.The system can show a typhoon path and intensity conveniently.