Research On Tropical Cyclone Climate Prediction Model Based On Feature Fusion

Tropical cyclone is a kind of extremely destructive meteorological disaster,often accompanied by wind,rain,storm surge and other phenomena.In recent years,global warming leads to the rise of ocean temperature,which strengthens the development and intensity of tropical cyclone and attracts people’s high attention.Located in the southern coastal region of China,Guangxi is affected by an average of 5 typhoons every year,which often cause serious economic losses and casualties.Therefore,improving the weather prediction ability of tropical cyclones is of great significance for preventing and reducing disaster losses in advance.In order to improve the prediction accuracy of the frequency and start date of tropical cyclones affecting Guangxi annually,this paper uses the sample data of tropical cyclones affecting Guangxi for 71 years from 1951 to 2021 provided by Shanghai Typhoon Research Institute,China Meteorological Administration,and 130 climate index data provided by the National Climate Center to establish a climate prediction model for tropical cyclones affecting Guangxi based on multiple linear regression,support vector regression,random forest and BP neural network methods,The ability to predict the climate of tropical cyclones affecting Guangxi has been improved,and the specific summary is as follows:(1)In response to the complexity of the physical factors affecting tropical cyclones,in order to obtain more comprehensive prediction factor information,a new technology for searching for factors affecting tropical cyclone frequency and start date is established,which combines machine learning and traditional statistical methods.Firstly,the frequency and start date of tropical cyclones were analyzed for correlation,and the high correlation factors were screened for the first time.The feature factors were selected using the stepwise regression and random forest methods.After two feature selections,redundant features were reduced and the generalization ability of the model was improved.(2)Establishing a tropical cyclone frequency and start date prediction model based on machine learning and feature fusion selection has improved the accuracy of tropical cyclone frequency and start date prediction.Establish support vector regression,BP neural network and random forest prediction models,and compare the prediction effects of models in different feature sets.The experimental results show that the selected impact factors have achieved good results in the four models,of which the random forest model has the best effect compared with the multiple linear regression,support vector regression and BP neural network methods.The average absolute error and average relative error are reduced by 0.75%,26.26%,0.16%,1.04%,and 0.37%,13.04%.The 8-year rating is 100 points,with an accuracy of 100%,providing new methods and ideas for predicting the frequency and start date of annual tropical cyclones.