Research On Typhoon Track Prediction Based On Deep Learning Method

Typhoon is a natural phenomenon,which can regulate the global heat distribution and maintain the heat balance.At the same time,the destructive power of typhoons is extremely strong.Every year,typhoon disasters frequently invade the coastal areas of the world,affect economy and endangering people’s lives.As an important function of the meteorological department,accurate typhoon track prediction can provide relevant meteorological departments with prior information,so that they can make more scientific and faster decisions and greatly reduce the losses caused by the typhoon.Therefore,it is particularly important for China and the world to improve the accuracy of typhoon track prediction.However,due to the complexity of typhoon movement and the scarcity of ocean observation stations,there are many limitations in using traditional methods to predict typhoon tracks.At present,typhoon track prediction mostly depends on numerical prediction.Its prediction results are relatively accurate,but the calculation is difficult and the cost is high.At the same time,researchers need to be very professional.So,this method is extremely demanding.In contrast,using machine learning(ML),especially deep learning(DL)to predict typhoon track,has the advantages of adaptive learning and nonlinear mapping and high fault-tolerance.The most important is that the deep learning method can greatly reduce the difficulty and cost of typhoon track prediction.In this paper,the deep learning method is used to predict typhoon tracks.In order to improve the prediction accuracy and reduce the cost,we construct a satellite image data set,and use the deep neural network predict the typhoon track.Finally,we use neural network to revise the prediction results to realize accurate and rapid prediction of typhoon tracks.The main contents of this paper are as follows:First of all,this paper describes the research background and significance of typhoon track prediction.We introduce the research status of mainstream typhoon track prediction at home and abroad,divides the current methods into two categories: numerical prediction and methods based on deep learning.We introduce the research results of two kinds of methods in typhoon track prediction,and focus on the advantages of deep learning method over numerical prediction.Then,several different deep learning technologies are introduced,including two neural networks in deep learning and some optimization strategies.Because there are few open-source satellite image data sets,this paper uses the satellite image of the Japan Meteorological Agency to build a data set.We mark the satellite images according to their relationship in the time dimension,and use the marked satellite images as the input of the prediction model.Then,based on the analysis of satellite images and their characteristics,a typhoon path prediction model is established based on CNN and LSTM.The hybrid dilated convolution is used in model to improve the problem of missing features of maker points,and the attention mechanism module is added to enhance feature extraction.The Deep Typhoon is built to predict typhoon track.Then,in order to further improve the prediction accuracy,based on the prediction results of the Deep Typhoon model,a grid-based coding method is proposed,which converts the prediction error vector into error coding.Based on LSTM network and the idea of multitask learning,TPEM model is constructed.TPEM model is used to predict the error,and the prediction error of typhoon track is corrected according to the prediction results,which further improves the accuracy of typhoon track prediction.At last,we conducted a comparative test,and the average absolute error of the Deep Typhoon model in the northwest Pacific and southwest Pacific test sets was 65.39 km and 62.41 km respectively.While ensuring the calculation speed and prediction cost,Deep Typhoon has achieved relatively accurate typhoon track prediction.On this basis,the TPEM is used to correct the prediction error.After modification,the prediction error of the Northwest Pacific data set is 63.94 km,and the prediction error of the Southwest Pacific data set is 59.18 km.Compared with the prediction results of Deep Typhoon,TPEM further improves the prediction accuracy of typhoon track.The results show that the Deep Typhoon combined with TPEM can accurately predict the typhoon track.Moreover,compared with the numerical prediction,the proposed method has faster prediction speed and lower prediction cost.