Statistical Post-processing Research On Probabilistic Forecasting Of Precipitation In Multi-model Ensembles

Although the ensemble forecast can greatly improve the prediction ability of uncertainty,the probabilistic forecast based on the ensemble system cannot fully describe all the uncertainties of the atmosphere.The statistical post-processing technique of ensemble forecast establishes a statistical relationship or statistical regression model between the model forecast data and the observed data by extracting the effective forecast information from each ensemble member.Thereby,it can correct the model prediction error and reduce the impact of model prediction uncertainty.Precipitation forecast is very important for weather forecast,forewarning of flood and debris flow.Because the observation sequence of precipitation(especially precipitation of the daily time scale and below)is discontinuous,the statistical post-processing of precipitation forecast is more difficulties and challenges compared with other meteorological elements.In order to solve the problem that BMAR is not suitable for post-processing of precipitation probability forecast and the problem that BMAs' EM algorithm can not get the global optimal solution,we proposed a L-BMAs model(the BMAs with limitedmemory BFGS).We use precipitation observations and forecasts(CMA,ECMWF,NCEP)on July 1 to August 31,2015,to build the L-BMAs model.And we test the precipitation probabilistic forecasts of the different ensemble system,which are after post-processing.We also compared the L-BMAs post-processing effect with BMAs post-processing effect.The main results as follow:(1)The post-processing test of precipitation probabilistic forecast of typhoon shows that the BMAs has an obvious improvement effect on the post-processing of ECMWF precipitation forecast.To some extent,it can reduce the false of the original precipitation forecast probability of typhoon rainstorm and above order.It can also improve the reliability of ECMWF precipitation forecast of each order.In addition,the 25th-95 th percentile prediction of BMAs can effectively improve the precipitation observation capture rate of typhoon Lekima.(2)The optimal training periods of BMAs and L-BMAs are different for different centers and different precipitation magnitude orders.In general,the optimal training period of CMA,ECMWF,NCEP and GE can be selected as 20 days,50 days,25 days and 35 days respectively.The optimal training period of L-BMAs can be selected as 45 days,20 days,25 days and 25 days respectively.When the training period is less than 25 days,the revised effect of BMAs on CMA is better than that of L-BMAs.When the training period is greater than or equal to 25 days,the revised effect of L-BMAs on CMA is better than that of BMAs.In addition,no matter how the training period is selected,the revised effect of L-BMAs on ECMWF is not as good as that of BMAs.And the revised effect of L-BMAs on NCEP prediction is always better than that of BMAs.(3)In the optimal training period,the CRPS's improvement after the BMAs postprocessed from high to low is multi-model ensemble(GE),NCEP,CMA and ECMWF.The ability to correct MAE from high to low is multi-model ensemble(GE),CMA,NCEP and ECMWF.In the optimal training period,the improvement degree of L-BMAs on CRPS from high to low is multi-model ensemble(GE),NCEP,CMA and ECMWF.And the correction ability of MAE from high to low is multi-model ensemble(GE),CMA,NCEP and ECMWF.Otherwise,the improvement of prediction effect is affected by the original prediction performance.The prediction effect of each ensemble system,whether improved by BMAs or LBMAs,from high to low is multi-model ensemble(GE),ECMWF,NCEP and CMA.(4)The comparison of the revised effect of L-BMAs and BMAs on different magnitude orders precipitation forecast shows that for trace precipitation,the ensemble forecast effect which revised by L-BMAs better than BMAs is ECMWF,NCEP and GE.For precipitation of light rain,NCEP and GE are the ensemble system with better prediction effect of L-BMAs.For precipitation of moderate rain,the ensemble prediction effect of L-BMAs better than BMAs is NCEP.For the precipitation above heavy rain,the ensemble forecast of the better prediction effect of L-BMAs than BMAs is CMA.