Using Global Ensemble Forecast System (GEFS) data, a regional ensemble forecast system based on WRF model is constructed. Two initialization schemes are tested to form the initial states of regional ensemble (namely DRUC ensemble). One is the direct dynamical downscaling of GEFS initial states (namely DOWN ensemble), and the other is the overlaying the downscaled initial perturbations of GEFS onto the analysis of high resolution regional numerical weather prediction (NWP) system, named Beijing Rapid Update Cycle (BJRUC) system. Using the two methods, a series of ensemble forecast tests are conducted, and the results show that the smallscale components of DRUC perturbations grow more rapidly than those of DOWN perturbations. For shortterm forecast, the DOWN perturbations tend to underestimate the forecast error while the DRUC perturbation tends to identify where the forecast error is large and where the forecast error is small. Ensemble forecast verification shows that the DRUC ensemble has larger spread and smaller root mean square error than DOWN ensemble at short forecast lead time, and the probabilistic scores of DRUC are also better for shortterm forecast. Typical precipitation case study shows that DRUC ensemble can provide better probabilistic precipitation forecast than DOWN ensemble.