One of the key factors by which to determine the quality of regional ensemble forecast is whether the initial perturbation of ensemble prediction can precisely reflect the structural characteristics of the forecast errors.In the present study,based on the data assimilation method,a new scheme of blending initial perturbation on the regional ensemble prediction system is proposed for the regional numerical prediction model of GRAPES.Specifically,the new scheme first introduces the global large-scale information as the background field and regional ensemble forecasts as observational data into the GRAPES m3DVAR system,then effectively integrates the high quality meso-scale information into global large-scale information,so as to construct a multi-scale initial perturbation of regional ensemble prediction.A case experiment and batch tests are carried out to compare the performance of the new scheme and original regional ensemble forecast.The results suggest that the multi-scale initial perturbation based on the data assimilation method is able to effectively combine the global large-scale information from the global ensemble forecast with the meso-scale information from the regional ensemble forecast,thereby leading to a more reliable probabilistic precipitation prediction.Therefore,this new scheme of blending initial perturbation of the regional ensemble prediction system is proven to be efficient in improving the quality of regional ensemble forecast,especially for the geopotential height field and temperature field,yet at the same time it has a slight effect on the wind field.