Abstract: The disparity between initial perturbations and forecast errors is a primary challenge in the Global/Regional Assimilation Prediction System–Regional Ensemble Prediction System (GRAPES–REPS). Herein, the Ensemble Transform Kalman Filter method is introduced to update the perturbations produced by dynamic downscaling in the GRAPES–REPS and can capture more mesoscale perturbations. Moreover, a new method called the cosine analysis constraint scheme is developed to optimize the updated initial perturbations, rendering the latest perturbations more consistent with the forecast errors. The results reveal that the perturbations adjusted using the cosine analysis constraint scheme may show the most abundant information on mesoscale perturbations compared to other experiments in the key area of precipitation. Notably, the structure and evolution of the perturbations become more reasonable and match better with the development of weather systems on different scales after constraining. In addition, the improvements in spread and perturbation energy are distinct, particularly in the early period of simulation.