In order to develop GRAPES (Global and Regional Assimilation and Prediction System) regional ensemble prediction system (GRAPES REPS), using ensemble transform Kalman filter (ETKF) as initial perturbation scheme, coupled with multiple physical processes combination as model perturbation, an regional ensemble prediction system is constructed based on the GRAPES_MesoV3.3.2.4 model in this paper. Besides, 40 d consecutive ensemble forecast experiments are conducted, the structure and evolution characters of ETKF generated initial perturbation are emphatically investigated, various methods are utilized to evaluate GRAPES REPS performance and its precipitation forecast capabilities, and a severe rainfall case is also analyzed to further illustrate the precipitation forecast performance of the EPS. The experimental results indicate that GRAPES REPS can generate promising initial perturbations characterized by flow dependent structure and good correspondence to the distribution of observation sites, and meanwhile the perturbations are orthogonal. Total energy of perturbations can keep appropriate growth in all forecast lead times. Ensemble forecast verification shows that ensemble forecast outperforms control forecast, the ensemble spread can maintain reasonable growth in 72 h forecast lead time. Comparisons between operational WRF REPS and GRAPES REPS on precipitation forecast are carried out, and the results show that GRAPES REPS outperforms WRF REPS. Case study indicates that ensemble forecast can provide much better heavy rainfall forecast than control forecast.