Revision And Interpretation Analysis Of Extended-period Precipitation Ensemble Forecasts By Machine Learning

In order to extend the applicability of historical reforecast data of numerical model in extended range ensemble forecast products,and to analyze the forecast skills of each member of extended range ensemble forecast,based on the ECMWF model data,this paper mainly studies the post-processing technology of the monthly precipitation forecast results of the extended range ensemble forecast in East China(23°N?38.5°N,113°E?123°E)in summer(June?October)with the forecast time of 30?45 days.??Statistical method is used to analyze the forecast members of historical reforecast and real-time extended range forecast,according to the evaluation of dispersion and root mean square error(RMSE),The results show that the application effect of ensemble forecast members and ensemble average forecast is analyzed,which provides a theoretical basis for constructing the feature engineering in this paper.??The historical deviation correction method is used to compare the extended range precipitation ensemble forecast data with the actual bilinear interpolation data,according to the mean absolute error(MAE)and system deviation(SB),the correction method is evaluated.The results show that in this paper,the error of historical deviation correction effect in each forecast time has not been significantly improved.??Based on AdaBoost?Bagging?Random Forest?Gradient Boost Regression Tree(GBRT)and Bayes Ridge Regression(BRR),the correction analysis of extended range precipitation ensemble forecast is carried out.The training was carried out by combining the ensemble average forecast and ensemble median data sets from several years(1999-2018)of historical reforecast data to construct feature engineering.The changes of mean absolute error(MAE)?system deviation(SB)and dispersion before and after the correction of extended range ensemble forecast are analyzed and compared from four aspects of forecast time?time series?spatial distribution and ratio value.The results show that:(1)In the 30?45 days forecast time limit,the average absolute errors and system deviations of the five regression prediction models corrected by using two dataset construction feature projects are better than those before ensemble forecast correction.Among them,using the ensemble average forecast dataset to construct the feature engineering regression model is better than the ensemble median forecast,and the AdaBoost algorithm is the most effective.(2)In the time series distribution,the errors of the five regression prediction models after the ensemble forecast correction from June to October fluctuate less and tend to be stable,and the average absolute error of the corrected results in July and August is significantly lower than that of the ensemble average forecast,and the correction effect is better.(3)In the error spatial distribution(excluding Taiwan Province and sea area),the average absolute error and systematic error of the five regression prediction models after correction are significantly reduced in the northern part of East China(31°N?38.5°N,113°E?123°E),especially in the plain area,and the correction effect of AdaBoost algorithm and Bayes Ridge Regression algorithm is better.(4)In the ratio value R of dispersion and root mean square error,the range of R corresponding to each forecast time is extended from [0.03,0.30] to [0.65,0.91] after the correction of five regression prediction models,and the fluctuation value is stable at about0.75,which enhances the rationality of ensemble forecast disturbance.(5)The historical reforecast data in numerical model forecast can improve the prediction skills for extended-range forecast.Based on five machine learning regression prediction models,this paper uses ensemble average forecast data set to construct a feature engineering correction method for the post-processing of extended range precipitation ensemble forecast,and the correction effect is better in the plain area in summer.