本文討論2003年梅雨季中尺度模式MM5 (Pennsylvania State University-National Center for Atmospheric Research Mesoscale Model, version 5)系集降水預報的成效，此系集系統包括有十七個MM5系集成員，由變動不同初始場、積雲參數法、微物理過程而組成。所有成員之系集平均降水預報在10mm以下降雨門檻的公正預兆得分（ETS）高於所有個別成員，顯示系集平均可以提升小雨降雨預報的技術，對於中、大雨則約位於中上的成績。 綜合平均的結果進一步發現，利用中央氣象局全球模式(GFS)為初始場相較於其他初始場可在中、大雨（降雨門檻10mm以上）獲得較佳的預報；在不同積雲參數法上，則以使用Grell積雲參數法者其降水預報為最佳；在微物理過程方面，小於10mm降雨門檻以使用Goddard方法者預測能力最佳，大於10mm門檻值以上則以使用Mixed-phase方法者為最佳。可見臺灣地區梅雨季之數值模擬比較適合使用來自GFS之初始場，搭配Grell積雲參數法與Mixed-phase或Goddard微物理方法之組合。 系集平均之降水預報，不論成員多寡，相對於單一模式成員皆有絕對之優勢；在初始場、積雲參數法、微物理過程三個變因之中，考慮所有變因所得之系集預報可以得到最佳的系集降水預報，考慮其中2個變因者次之，而僅考慮1個變因所得之系集預報則為最差；此三個變因中又以變動初始場最有利於系集降水預報，變動積雲參數法次之，變動微物過程最差。

This paper presents the verification results of MM5 ensemble precipitation forecasts during the 2003 Mei-yu season. The ensemble included 17 members that varied from 3 different initial conditions (ICs), 2 cumulus parameterization schemes (CPSs), and 4 microphysics schemes (MSs). The ETS (equitable threat score) of the ensemble mean of all members were the highest at rainfall thresholds smaller than 10mm, and above the average at larger thresholds among all members. From the results of different ensemble means among the members, it is found that simulations using the GFS IC could obtain the best MM5 rainfall forecasts among the 3 available ICs; runs with the Grell CPS presented better ETS than the Kain-Fritsch CPS; members using mixed-phase and Goddard MS provided better rainfall forecasts than the other 2 MSs. It is therefore recommended that when running MM5 in the Taiwan are for a Mei-yu season, the best setting should include the Grell CPS and the mixed-phase or Goodard MS, and using the GFS IC as initial fields. It is also found that the ensemble mean precipitation forecast has advantage against forecasts from a single member. The ensemble that varies from all three controls (IC, CPS, and MS) has the best rainfall forecasts, those varying from two out of the three available controls perform the second, and those varying from only one control have the lowest ETS. Among these, varying IC is the best, CPS the second, and MS the worst, in making an ensemble for producing ensemble rainfall forecasts.