Research On Typhoon Movement Forecasting In The Northwest Pacific Based On Deep Learning

The Northwest Pacific Ocean is one of the most active typhoon regions in the world.China is affected by landfall typhoons every year,causing catastrophic weather such as gale-force winds,heavy rains,and storm surges in coastal areas.Therefore,accurate prediction of typhoon paths in the northwest Pacific region is very meaningful for disaster prevention and mitigation in China.In this thesis,we propose to use the best track data of China Typhoon Network(CMA-BST)and ERA5 reanalysis data fused to form multivariate time series data to study the method of predicting the path of typhoons in the northwest Pacific Ocean by a recurrent neural network,and the main work is as follows:1)Using the CMA-BST data from 1959 to 2020,we statistically analyze the characteristics of typhoons in the Northwest Pacific region and find that the number of typhoon occurrences in the Northwest Pacific gradually decreases,but the proportion of super typhoons is increasing.The K-mean clustering algorithm is used to analyze the paths of typhoons in the northwest Pacific.The main paths have the following four categories: the first category of westward-turning,the second category of west-northwestward-turning,the third category of eastward-turning,and the fourth category of west-northwestward-turning offshore.All four types of typhoons are located at low latitudes of 5-10 degrees north latitude,and all gradually die out near land areas or at high latitudes and high longitudes of the ocean.The directional paths are more often than linear paths and are mostly northeast and northwest oriented.2)To address the problem of poor quality of typhoon datasets and the inability of the datasets to integrate different kinds of information.In this thesis,firstly,the record points with weak TC and the record points with a sampling frequency of 3 hours are removed from the CMA-BST data to improve the quality of the CMA-BST data.The nearest neighbor interpolation method is used to make the spatial resolution of CMA-BST data and ERA5 data consistent.The data in the range of 10E10 N around the corresponding TC center are extracted,and after averaging the values,they are concatenated with the CMA-BST data information to form multivariate time series data.3)Using the established time series data,three recurrent neural network models,RNN,LSTM,and GRU,are used to predict the 6h,12 h,and 24 h paths of the typhoon.The experiments show that recurrent neural networks can make good use of the time series of data,and the predicted paths are feasible.The longer the prediction time of the three models,the worse the prediction effect,among which LSTM has the best effect for long prediction time,the path prediction error is less than 173.15 km,but there is still a gap compared with the accuracy of other artificial intelligence forecasting methods;in addition,the prediction accuracy of the three models in the longitude direction is much lower than that in the latitude direction,and the prediction effect is poor for special areas such as high latitude and land.4)According to the shortcomings of the three prediction models,the article studies the improvement of LSTM,mainly by introducing the attention mechanism and Encoder-Decoder method,and develops the improved AM-Bi LSTM model.The curvature feature of the moving path is added to complement the deflection and steering information to solve the problem that the error in the longitude direction is much larger than that in the latitude direction.Using the Bi LSTM as an encoder and decoder,it can effectively utilize the typhoon path timing information and improve the accuracy of high latitude and post-landfall predictions.Experiments show that the prediction accuracy of 24 h using this model can reach 153.06 km,which effectively improves the prediction accuracy of the typhoon path for 24 h.