Abstract: To alleviate the problem caused by fixed boundary condition in the GRAPES-RAFS system and improve the forecast qualities of regional model, a blending method to merge large scale information of global model with meso-scale information of regional analysis is implemented in the GRAPES-RAFS system using Discrete Cosine Transform (DCT) filter. During 9 May to 9 June 2018, the forecast qualities between the control experiments of the cold-start and experiments of warm-up initialization and the blending scale analysis by the GRAPES-RAFS are compared. The numerical results indicate that the large scale information and variables′ harmony of analyses increase, and the spin-up time is decreased with the global information and analysis in GRAPES-RAFS at 3 h intervals merged. The forecast quality of surface variables, analysis and precipitation with warm-start has been obviously improved. In addition, the simulated track of typhoon Ewiniar is much closer to observation by using the blending analysis. The warm-up initial physical fields, which lead to far smaller surface pressure tendency, are more harmonic than those of cold-start initialization. However, it is observed that the warm-up forecast quality of 2 m temperature and 10 m wind is not better than that of the cold-start, and that the warm-up precipitation score of ETS is lower than that of the cold-start as well.