According to the normal distribution characteristics of initial atmospheric errors,in order to describe the uncertainties in the initial conditions of the GRAPES global model,this paper used the Gausssian sampling method to construct initial perturbations for GRAPES global ensemble prediction system (GRAPE-GEPS) from GRAPES singular vectors (SVs),which are specifically designed for ensemble forecasting.Using the global assimilation analysis field of GRAPES,this paper carried out the ensemble experiments of GRAPES-GEPS with SV-based initial perturbations.Results show that the growths of ensemble spreads of geopotential height,wind speed and temperature increase quickly,and the growth rate remains stable during the forecast range.The small distance between ensemble mean RMSE and ensemble spread indicates that there are reasonable spread-skill relations in the GRAPES-GEPS.Compared with RMSE of unperturbed control forecast,there are clear advantages observed in RMSE of ensemble mean,particularly in the middle range forecasts.The stochastically perturbed parametrization tendency (SPPT) scheme is used to represent the model uncertainties in the GRAPES-GEPS.It can be seen that the inclusion of the SPPT scheme improves the ensemble spreads of geopotential height,temperature and wind speed at each forecast lead time in the Northern (Southern) Hemispheric extra-tropics and the tropics in particular,and improves the spread-skill relations as well.The SV-based initial perturbations and SPPT scheme have been used in the operational GRAPES-GEPS at China Meteorological Administration since December 2018.