Research On Short-term Climate Prediction In China Based On Machine Learning Method

Short-term climate prediction has extensive social needs,but due to the complexity of the formation of the climate system and the uniqueness of China’s topography,it is extremely complex and difficult to accurately predict it.At present,the main forecasting methods are based on dynamic models and statistical forecasting methods.However,the dynamic mode is sensitive to calculating the initial value and cannot assimilate long-term historical data;The statistical prediction method cannot solve the problems of model instability,feature factor complexity,and nonlinearity of the forecast factor-forecast quantity relationship,so the actual forecasting skills of these two methods are not high.Machine learning is also inherently a statistical method,which can solve the problem of model nonlinearity,but it is not widely used in short-term climate prediction.Based on the above reasons,in order to deal with the nonlinear problem of forecast factor-forecast quantity relationship,this study proposes a machine learning prediction model and predicts China’s average monthly temperature and precipitation.Aiming at the problems of model instability and complexity of feature factors,this study proposes a time difference machine learning model(TD-ML)scheme to predict seasonal temperature and precipitation in China.Compared with statistical prediction methods and current climate dynamic prediction models,the prediction model proposed in this study shows great advantages in challenging prediction techniques for the spatial distribution of climate anomalies.The specific research content is as follows:In view of the non-linear characteristics of the relationship between the forecast factor and the forecast quantity,using the monthly climate data of 160 national stations in China and the130 climate index sets of the National Climate Center,the Pearson correlation analysis was used to determine the forecast factor,and four algorithms,namely,linear regression,ridge regression,random forest,and gradient lifting tree algorithm,were used to establish a set of prediction scheme models for China’s monthly temperature and precipitation,The prediction results are compared with those of three international mainstream dynamic models to test the prediction effect.The results show that the prediction ability of machine learning algorithm model is slightly improved compared with the existing dynamic model.In view of the two statistical prediction difficulties of seasonal scale temperature prediction,which are model instability and feature factor complexity,the dynamic modeling and feature enhancement methods are proposed respectively,and a new dynamic modeling machine learning prediction scheme is formed on this basis.The results show that both dynamic modeling and feature enhancement can improve the skills of machine learning models,so that the final TD-ML scheme can be better or not weaker than them in trend control when compared with the same period prediction of international mainstream business models,and has greater advantages in abnormal spatial distribution skills.This paper probes into the reasons why the seasonal precipitation prediction effect is not ideal,and takes summer precipitation as an example to analyze the impact of different reporting months on the accuracy of seasonal precipitation prediction in China.The experiment shows that the starting month including the meteorological factors of the previous winter season can effectively improve the prediction accuracy of the model for summer precipitation.