MM5 ensemble mean forecasts in the Taiwan area for the 2003 mei-yu season

This paper presents an evaluation study of a real-time fifth-generation Pennsylvania State University-NCAR Mesoscale Model (MM5) mesoscale ensemble prediction system in the Taiwan area during the 2003 mei-yu season. The ensemble system consists of 16 members that used the sam nested domains of 45- and 15-km resolutions, but different model settings of the initial conditions (ICs), the cumulus parameterization scheme (CPS), and the microphysics scheme (MS). Verification of geopotential height, temperature, relative humidity, and winds in the 15-km grid shows that the members using the Kain-Fritsch CPS performed better than those using the Grell CPS, and those using the Central Weather Bureau (CWB) Nonhydrostatic Forecast System (NFS) ICs fared better than those using the CWB Global Forecast System (GFS) ICs. The members applying the mixed-phase MS generally exhibited the smallest errors among the four MSs. Precipitation verification shows that the members using the Grell CPS, in general, had higher equitable threat scores (ETSs) than those using the Kain-Fritsch CPS, that the members with the GFS ICs performed better than with the NFS ICs, and that the mixed-phase and Goddard MSs gave relatively high ETSs in the rainfall simulation. The bias scores show that, overall, all 16 members underforecasted rainfall. Comparisons of the ensemble means show that, on average, an ensemble mean, no matter how many members it contains, can produce better forecasts than an individual member. Among the three possible elements (IC, CPS, and MS) that can be varied to compose an ensemble, the ensemble that contains members with all three elements varying performed the best, while that with two elements varying was second best, and that with only one varying was the worst. Furthermore, the first choice for composing an ensemble is to use perturbed ICs, followed by the perturbed CPS, and then the perturbed MS.