This paper reports the first effort on real-time ensemble predictions of precipitation during the 2000--2002 Mei-Yu seasons (May to June) over the Taiwan area. Six members were included, each using the fifth-generation Pennsylvania State University-National Center for Atmospheric Research Mesoscale Model (MM5) nesting down to 15-km grid size, with different combinations of cumulus and microphysics parameterizations. Rainfall forecasts were evaluated with the equitable threat score (ETS) and bias score (BS). On the basis of verifications on 15-km grid points over three Mei-Yu seasons, it was found that no one member persistently had the least root mean square error of 12--24 hours and 24--36 hours accumulated rainfalls. For rainfall occurrence, most members had better predictions over the northeastern mountainous area, the northwestern coastal plain, the central mountain slope, the southwestern coastal plan, and the southwestern mountainous area. These regions also corresponded to areas of more accumulated rainfalls during three Mei-Yu seasons. An ensemble prediction, using a multiple linear regression (MLR) method which performed a least-square fit between the predicted and observed rainfalls in postseason analysis, had the best ETS and BS skill. The MLR ensemble forecast outperformed the average forecast (for all six members), the average forecasts of cumulus (four-member) and microphysics (three-member) ensembles, and also a high-resolution (5-km) forecast; however, a high-resolution forecast still had better skill for heavy rainfall events. The MLR ensemble forecast, using the weightings determined from previous Mei-Yu seasons, still had similar ETS trend to that with weightings determined by current-year Mei-Yu season, albeit with less skill.