Research On Ensemble Forecast Precipitation Deviation Correction And Probability Forecast Method

Precipitation is formed under the influence of a series of complex physical processes and is influenced by various weather systems.Therefore,the prediction accuracy of precipitation is relatively lower than that of weather elements such as air pressure,wind speed,and temperature.It has become one of the most difficult weather elements in forecasting and climate prediction.Because precipitation is the result of interactions between different weather systems,its spatial and temporal distribution is complex.The statistical characteristics are different on different time scales,especially the precipitation in the daily precipitation or the shorter time scale is skewed distribution.So the use of precipitation probability prediction is better.Although numerical prediction has made great progress,the errors resulting from data analysis and assimilation processing will lead to uncertainty of the initial field of numerical prediction model.The ensemble forecast is considered as one of the most effective ways to get the probability forecast.The precipitation probability prediction produced by ensemble prediction products has improved the accuracy of precipitation forecasting.But the uncertainty of precipitation and its numerical model,there is still errors in ensemble precipitation prediction.Therefore,it is a new trend in the application of ensemble forecasts to evaluate and calibrate the ensemble precipitation prediction in recent yearsIn order to calibrate the ensemble forecasts and study the probability distribution and ensemble precipitation forecasts error.First,generalized additive model is used to analyze the relationship between prediction center,prediction time and model resolution on the ensemble precipitation forecasts.Then,precipitation ensemble forecasts from three global numerical forecast centers(CMA,ECMWF and NCEP)and surface precipitation observations in southeast China are used to study the precipitation deterministic error,discrete error and probability error under different prediction time.Based on Bayesian theory,historical observations and prediction data are provided prior information and establishing the Bayesian probability precipitation forecasting model for Chengdu and Guangzhou station to analyze the correction with different schemes.Finally,based on the 24 h precipitation ensemble forecasts from three global numerical forecasts centers and surface precipitation observations in southeast China from 1 June to 31 July 2015 and using the Bayesian Model Averaging(scheme A)and combined Bayesian Model Averaging and statistically downscale(scheme B),we have corrected the three single-center and multicenter grand ensemble precipitation predictions,compared the adjustment effects of the two schemes,and the selected the precipitation forecasts from August 1 to 31 in 2015 to perform an independent sample test,and analyzed the skills of precipitation prediction before and after the correction.The main results are as follows:(1)Different prediction center models,models resolution,and forecasting time are used as explanatory variables.Based on the generalized additive model,analyzing the relationship between them and precipitation prediction.The results show that forecasting time has the most significant impact on precipitation ensemble forecast,but the influence of model resolution from different centers on precipitation forecasting is also important.(2)Characteristic analysis of deterministic prediction error of precipitation ensemble forecast shows that the large value of root mean square error is near the southeast coast and the error and range of the root mean square error are gradually increasing with the increase of the forecast time.TS score of ECMWF to small scale precipitation is high.CMA has a high TS score to large scale precipitation,but there are a large of false alarms.The comprehensive comparison shows that the prediction quality of ECMWF is the best.(3)The dispersion between members of the three operational centers is not enough,observation always fall outside range between maximum and minimum values of ensemble members.According to the Rank distribution,the ensemble forecast system estimates the precipitation in the coastal areas is small.(4)The BS score of ECMWF is the smallest,so the prediction quality of ECMWF is the best.With the increase of the forecasting time and precipitation threshold,the BS score gradually decreased.The main reason for this phenomenon is that the precipitation probability forecast of the 0.1mm threshold has a lot of false alarms,and the precipitation threshold is 50.0mm,there are less 50.0mm stations and less time to predict and observe.In addition,the BS score in the coastal areas is large,indicating that the center has a poor forecast effect on large precipitation in the coastal areas.(5)The southeast China is the “scale parameter-dominated” region and the coastal areas have a high probability of heavy rain.And the distribution of shape parameters and scale parameters has high correlation.By the information score of Bayesian precipitation probability forecasting model shows that BMA has the most significant effect on the Bayesian precipitation probability prediction based on the observed prior information and the grand ensemble prediction.(6)Taking 50 th percentile precipitation forecast of three single center and grand ensemble as example,the results indicate that scheme A eliminates a large number of the false alarm of light rain and corrects the bias for light and moderate rains remarkably.But the correction of the precipitation intensity for those exceeding heavy rain is not evident.Scheme B not only reduces the false alarm of raw ensemble forecasts,but also corrects the precipitation intensity and the rainfall area.So that the range and magnitude of the precipitation forecasts are closer to observations,the best correction for the GE.(7)However,the precipitation forecasts calibrated by B scheme remain false alarms and missing,especially for the high rainfall and extreme precipitation.And area's shape or trend in little change before and after correction.This shows that method is not obvious to improve the fall of rain and extreme precipitation forecast.The result is closely related with the raw ensemble prediction,the improvement more depend on own ability or physical process to eliminate the raw ensemble prediction system error.