Application Research Of Machine Learning Method In Nowcasting Forecasting

Traditional approaches to nowcasting forecasting are mainly based on numerical methods,extrapolation techniques and conceptual models.There are many factors to be considered in these methods,the construction of the model is too complex,the forecasting steps are too many,and the cumulative error is large.In recent years,machine learning methods have been widely used in weather forecasting.In this paper,three machine learning methods,Support Vector Machine(SVM),Gradient Boost Decision Tree(GBDT)and eXtreme Gradient Boosting(XGBoost),are introduced to predict the nowcasting forecasting base on radar rainfall products.Threat score(TS),Probability of Detection(POD),False Alarm Ratio(FAR)and Miss Alarm Rate(MAR)are used to synthetically compare the effects of three machine learning methods on the approaching rainfall prediction.XGBoost method and PPLK method are combined to forecast the short-term Quantitative Precipitation Nowcasting,and the relevant test indicators are used to evaluate it.The following are the main results and conclusions of this paper:(1)The nowcasting forecasting method based on SVM is proposed.The sunshine and rain forecast model base on non-linear support vector machine with radar rainfall products is established,which considering less influencing factors.The probability of detection in radar rainfall product with more concentrated rainfall distribution is higher.The feasibility of this method is verified by experiments.(2)The nowcasting forecasting method based on GBDT is presented.The nowcasting forecasting model base on Gradient Boosting Decision Tree with radar rainfall products is established,and the model is data-driven.The approach with GBDT for nowcasting forecasting is good in all kinds of radar rainfall products.The experiment proves that the method is feasible.(3)The nowcasting forecasting method based on XGBoost is proposed.The XGBoost model of radar rainfall product-weather forecast is established.Experiments show that the XGBoost model has a strong ability on nowcasting forecasting in different radar rainfall products.The Threat Score and Probability of Detection of the overall forecast are at a high level.Experiments verify the validity and accuracy of the method.(4)After comparing the results of XGBoost method with those of SVM method,RF method and GBDT method,the experimental results show that the prediction performance of XGBoost method is better than the other three methods,and the prediction accuracy is the highest.(5)The combination of XGBoost method and PPLK method is used for short-term Quantitative Precipitation Nowcasting forecasting.Experiments verify that the combination of XGBoost method and PPLK method has better prediction performance than PPLK method.It also confirms the practicability and efficiency of XGBoost method in rainfall prediction.