本文利用12組不同物理組合的WRF(Weather Research and Forecasting Model)模式系集成員，針對2004年5月15日~6月15日臺灣地區梅雨季進行一整個月的連續積分，再將降雨產品分別和綜觀測站觀測、臺灣自動雨量站觀測進行定量降水預報(Quantitative Precipitation Forecast , QPF)校驗，藉以評估WRF模式降水預報能力，並找出適合臺灣地區及華南地區梅雨季降水模擬的最佳物理組合。從降水校驗的公正預兆得分(ETS)分析得知WRF模式降水模擬在12~24小時有最佳表現，小雨模擬能力比大雨強，而且使用Yonsei University (YSU)邊界層參數法比Mellor-Yamada-Janjic (MYJ)邊界層參數法可得到較好的降水模擬。在華南地區採用WRF Single-Moment (0WSM 5-class微物理法搭配Kain-Fritsch積雲參數法可有最好的結果，在臺灣地區的小雨模擬也是採用此物理組合最好，但是臺灣地區大雨模擬較適合採用表現穩定的Grell-Devenyi積雲參數法，不過Grell-Devenyi積雲參數法有高估大雨的缺點。此外，模式系集平均的降水預報能力不錯，大多排名在1~3名之間。

This paper presents a quantitative precipitation forecast (QPF) verification study of the Weather Research and Forecasting (WRF) model during the 2004 Mei-yu season. Twelve members of WRF simulations, with different physics combinations, were run twice daily from 15 May to 15 June 2004. The 12-h accumulated precipitation forecasts from the 15-km grid were verified suing rainfall observations in Taiwan and Southeast China. The equitable threat score (ETS) show3s that the precipitation forecasts between 12 and 24 h of simulation had the highest scores among other time periods and the scores decreased as the thresholds increased. The model with the Yonsei University (YSU) PBL performed better than that with the Mellor-Yamada-Janjic (MYJ) scheme. In southeast China, the combination of the WRF Single-Moment (WSM) 5-class microphysics scheme and the Kain-Fritsch cumulus parameterization scheme was the best among others. In Taiwan, this physics combination also performed the best at small rainfall thresholds, but at large threshold the Grell-devenyi cumulus parameterization scheme is recommended. The weakness of this scheme, however, is its tendency of over-prediction. Finally, the ensemble mean issued consistently good rainfall forecasts, usually above the top 30% among all single members.