Research On Tropical Cyclone Track Prediction Method Based On Multi-modal Data

Tropical cyclone(i.e.typhoon or hurricane)has a negative impact on human life and natural environment for its high occurrence frequency,great harm,wide impact range and long disaster chain.The research on tropical cyclone and its characteristics has been paid much attention by many experts and scholars,and tropical cyclone track research is one of the key topics in this field.Traditional tropical cyclone track prediction methods are mainly divided into dynamic method(also known as numerical prediction model)and statistical model.Although it has achieved a good performance to a certain extent.However,there are some disadvantages: the dynamic method relies on strong computational power and the ability to model the weather state dynamically.And traditional statistical methods often rely on manual feature extraction,which is inefficient and relies on domain knowledge of experts.In addition,under the future climate conditions of global warming,severe tropical cyclones have many trends year by year.And more and more factors affect the cyclone,which also increases the difficulty of trajectory prediction.The current bottleneck of tropical cyclone track prediction lies in :(1)It is difficult to solve the long-term prediction problem,lead to poor long-term forecast performance(2)the lack of considering the influence of the external environment factors for trajectory,it is difficult to extract spatial and temporal correlation between meteorological factors and trajectory characteristics,cause problems for extreme path prediction(3)It fails to make full use of various types of data and realize value complementarity among different modes,which leads to the bottleneck of overall prediction efficiency and accuracy.In recent years,deep learning method has been proved to be able to automatically extract spatial and temporal features from large scale data,and it is highly efficient and accurate for data prediction of complex structures.The deep learning method provides new solutions and ideas for the bottleneck of tropical cyclone track prediction.Based on this,this paper constructs a novel ensemble neural network trajectory prediction model based on multi-modal data.To some extent,it can solve the problems including the difficulty of long-term prediction and the poor performance of extreme path prediction.(1)Aiming at the cyclone's attributes and meteorological factors in the numerical Best Track Data,a new prediction model based on the numerical data was constructed.First,autocorrelation and partial autocorrelation analysis are used to determine the best input history time.Then the model based on sequence-to-sequence(Seq2Seq)neural network architecture with attention mechanism and noise auto-encoder(DAE)network is proposed,in which DAE is used for feature reconstruction and associated feature extraction.Seq2 Seq is used for long-term temporal feature extraction.(2)For satellite remote sensing images,a new prediction model based on image data is constructed.Firstly,it is difficult to map from image data to numerical data,Predicted target(latitude and longitude)is converted into density map.Then,on the basis of fully considering the influence of cloud wall area and spiral rainband area outside typhoon eye on the trajectory,the Conv LSTM based on spatial attention mechanism was used to extract the global and local spatial and temporal features.The Encoding-forecasting framework is used to accurately predict the track of extreme cyclones.(3)Multiple modal data are fully used to make up for the deviation of results caused by a single data.In this paper,a new ensemble method based spatial-temporal trajectory is proposed,which integrates the pre-trained prediction network based on numerical data and image data.For realizing the accurate prediction of the track of tropical cyclones with long span time range and extreme trajectory.