Numerical weather prediction (NWP) models are commonly used for wind power forecasts, but NWP forecasts are uncertain due to uncertainties in the initial conditions, approximate model physics, and the chaotic nature of the atmosphere. Ensemble prediction systems (EPS), which simulate multiple possible futures, thus provide valuable information about forecast uncertainties. However, the spatial resolution of global ensemble forecasts from the European Centre for Medium-range Weather Forecast (ECMWF) and the National Centers for Environmental Prediction (NCEP) is relatively coarse and insufficient for many wind power farms built in complex terrain. This work proposes using the Weather and Research Forecasting model (WRF) to downscale ECMWF EPS and NCEP global ensemble forecast system (GEFS) to determine and compare the added values of downscaling different global EPS forecasts for wind forecasts in the complex terrain of Sichuan and Yunnan in China. A total of 366 days of day-ahead forecasts (28 to 51 hours) for wind speed at 80 meters are evaluated. The results demonstrate that the ensemble average of the higher resolution WRF downscaled forecast is considerably better than that of the global EPS forecast, and downscaled forecast of ECMWF EPS achieves the best performance. Also, a selective ensemble average (SEA) method is proposed and applied for the ultra-short (10 to 13 hours) forecast. Verification results demonstrate that the SEA method outperforms the ensemble mean.