Abstract: To improve short-range weather forecast predictability of high impact weather process, a set of national-level mesoscale ensemble prediction system (MEPS) is developed at National Meteorological Center (NMC). The theory to set up an ensemble prediction system lies in the following facts: There are no perfect forecast models, and atmosphere is a chaotic dynamical system, so any small error in the initial condition will lead to growing errors in the forecast, eventually leading to a total loss of any predictive information. The MEPS at NMC takes advantage of achievements at high resolution deterministic mesoscale prediction model, data assimilation system as well as experience from development of global ensemble prediction system. The error growth features for mesoscale model forecast within China area is explored and it is found that most of the convective-scale weather system develops in weak baroclinic environment and the quick growth errors resulted from baroclinic instability. Considering characteristics of the circulation regime, season and geographical domain, the initial perturbation technique of breeding method is adopted to perturb the initial fields. Furthermore, to reflect uncertainties within physical process as well as systematic errors within mesoscale model, many options of microphysics, convective cumulus parameterization, boundary layer schemes, land surface process schemes and combinations in the model are tested for a certain period to evaluate the performance of different schemes. The experiment indicates that physical process perturbation has equal or even greater impacts on spread of ensemble prediction comparing with initial condition perturbation. Therefore, assembling of different microphysics schemes, cumulus parameterization, and planetary boundary layer processes is applied to build a multi-initial condition, multi physics ensemble system. The initial conditions and lateral boundary conditions are obtained from global ensemble system at NMC and trickily rescaled during model integration process. To reduce systematic bias in ensemble forecasts, an adaptive Kalman Filtering algorithm is applied as bias correction method and the results is inspiring. Ensemble forecast products include ensemble averages, spread and probability of multiple elements (wind, temperature, humidity, geopotential height, rainfall, etc.) in multiple layers are produced and performance of the ensemble system is evaluated. To evaluate the performance of NMC's regional mesoscale ensemble prediction system, different ensemble verification methods is used to estimate and compare 6 mesoscale ensemble prediction systems within a common forecasting configuration during the WMO/WWRP Beijing 2008 Olympics Mesoscale Ensemble Research and Development Project. Results indicate that the overall predictability of mesoscale ensemble prediction system at NMC is overall comparable to the international participants. The MEPS is still not good enough for fixed site and time-specific forecasts, but it demonstrates good ability to capture the high impact weather event and will play a role in everyday forecast.