A Comparative Study Of Several Algorithms For Typhoon Prediction In The South China Sea Based On Machine Learning

The South China Sea is one of the regions with high typhoon prevalence.The ravages of typhoons in the South China Sea have brought incalculable losses to the residents of China’s southeast coast.Accurate prediction of typhoon paths is of great significance for disaster prevention and avoidance.This paper uses the typhoon dataset in the South China Sea from July to September as an example to explore how to improve the accuracy of typhoon path prediction.The influencing factors of typhoon path are complex,and it is difficult to accurately predict the typhoon prediction of different path characteristics by a single model.The feature extraction of forecast factors is one of the important factors affecting the prediction results.Therefore,in view of several difficult problems affecting typhoon path prediction,three corresponding prediction methods are proposed.The main research work is as follows:1.Aiming at the limitation of a single model to predict typhoons with different path changes,this paper proposes a probabilistic prediction model based on MDN.Using MDN to combine multiple probability functions,the probability distribution of predicted values can be better fitted.And the probability distribution map of typhoon path is given by the combination of probability functions of longitude and latitude,which is convenient for application in typhoon forecasting.2.Aiming at the problem that the feature extraction and learning of predictors are inadequate,this paper proposes a prediction model based on CGAN and LSTM,considering the excellent learning ability of CGAN.The model structure of CGAN is modified by LSTM and MLP.As a result,various features of predictors can be extracted respectively.And the loss function of this model is improved.3.Aiming at the problem that the deep network is easy to lose the basic features of the predictors,this paper proposes a Bagging ensemble prediction model based on BLS.The mapping nodes of BLS can preserve the original information of predictors as much as possible,and the ensemble way of Bagging can help to improve the prediction accuracy of BLS as a basic model.In this paper,the stepwise regression model is used as the benchmark model.And comparative experiments were carried out on the three proposed prediction models.The experimental results show that the three models proposed in this paper reduce the prediction errors of 16.28 km,8.76 km and 22.82 km,respectively.Finally,this paper compares and analyzes the performance,advantages and disadvantages of the three models.