Time-Lagged Ensemble Methods Applied For Extended-Rang Forecast Of Precipitation By Using Atmospheric General Circulation Model

In the background of global warming, extreme weather and climate eventsfrequently occur. Except weather forecasting in one week andsub-seasonal/seasonal climate prediction, researching and developing the newnumerical forecasting technique for6-30day extended-range prediction is alsovery significant to addressing these extreme events. Based on the Beijing ClimateCenter’s Atmospheric General Circulation Model, the main purpose of this paperis to study the forecast method of6-30day extended-range precipitation. First,we conducted a relatively spread time-lagged ensemble system by using a largenumber of hindcasts during1996-2005summers, and evaluated the equal weightensemble mean method (LAF) used to6-15day forecasts in terms of EquitableThreat Score (ETS), Frequency bias (BIA), Hanssen and Kuipers score (HK) andPropotion Correct (PC). It indicates that this method can improve thedeterministic forecast. Moreover, we suggested a precipitation categoricalensemble forecast method (LCF). It further shows a promise improvement for6-15day summer precipitation in China as compared to the ensemble mean. Lastbut not least, we studied predictability of the occurrence frequency of dailyprecipitation during every5days in30days, and the result indicates that the LCFcan significantly enhance this occurrence frequency prediction. In addition, wediscussed the influence of ensemble forecast from different resolutions in thispaper. The main conclutions as following:(1) Through examining the LAF predictability of500hPa geopotential heightin the globe, we found that time-lagged forecasts within the last three days fromBCC_AGCM hindcasts could generally contain useful initial information in theextended-range. Based upon this result, the different time lag interval sensitiveexperiments show that the LAF using four members at24hours time lag intervalsand last five members at12(or6hours) time lag intervals can significantlyimprove the6-15day precipitation for1mm and5mm thresholds, comparatively,the LAF using6hours time interval members is optimal (i.e. the optimal LAF); (2) As ensemble probabilistic forecasts get much more divergence withlonger lead time, we suggested a time-lagged categorical precipitation forecast(LCF) method. It is not a probability forecast but a categorical forecast ofprecipitation intensity at any model grid box. A given categorical precipitation isforecasted only when the ensemble probability for that categorical precipitationto occur at one grid box exceeds a certain threshold. Based upon13memberswith lagging3days at6hours intervals, the ETS, HK, BIA and PC scoreevaluations for the LCF with different probabilistic thresholds indicate that theLCF with5/13and4/13thresholds can most significantly improve the6-15day1mm and5mm categorical precipitation forecasts as compared with thedeterministic forecast and the optimal LAF, especially for the1mm rainfall andabove (1+mm), the correspondeing ETS and HK scores respectively increaseabove0.1and0.2, the PC score is higher than60%, and the frequency bias BIAscore is closer to1.0than the LAF. In addition, similar improvements by LCF arealso found for the prediction of several other categories of precipitation, and theoptimal threshold for LCF method to achieve the best results slightly increases asthe rainfall threshold decreases.(3) The results of geographical distribution of6-15day forecasts accuracyrates further show that the improvements by LCF are primarily located over therainy regions where the frequencies of observed1+mm rainfall days during theforecast time period are larger than about40%-50%. These regions mainlyinclude central to southern, northeastern China and the northeastern part of theTibetan Plateau, the corresponding PC scores increase about5%-15%ascompared to the deterministic forecast. The geographical distribution ofdifferences of PC score between the optimal LCF and the optimal LAF indicatethat the LCF is generally better than the LAF, and the accuracy rates increaseabout3%-6%. Meanwhile, the LCF can decrease the LAF flase alarms in thepart of arid and semi-arid drought regions over China.(4) In this work, the LCF method is also used to predict the occurrencefrequency of daily precipitation during every5days. The results show that thisfrequency larger than1day,2and3days during every5days is predictable byusing the LCF, especially the regions where the frequencies of1+mm/dayobserved rainfall days are larger than50%, the corresponding PC values are70%-80%,60%-70%and50%-60%, respectively. (5) The comparation of ensembles respectively based on BCC_AGCM at T42resolution and T106resolution shows that LAF and LCF are also useful toimprove6-15day1+mm and5+mm daily precipitation forecasts at a highermodel resolution, and the optimal probabilistic threshold is not sensitive to themodel horizontal resolution. The occurrence frequency of1+mm dailyprecipitation during every5days is also predictable by using the optimal LCF.