The horizontal diffusion term is often used in the model to suppress nonlinear calculation instability or damping of false short waves, but this will cause excessive dissipation of small-scale kinetic energy in the numerical model near the truncation scale, therefore causing prediction uncertainty. In order to compensate the excessively dissipated small-scale kinetic energy back to the model, a Stochastic Kinetic Energy Backscatter (SKEB) method is introduced into the CMA regional ensemble prediction system. First, based on the first-order autoregressive random process in the horizontal direction to expand the spherical harmonic function to obtain the random field, and then construct the random flow function forcing, transform it into horizontal wind speed disturbance, and make random compensation for the dissipated kinetic energy. We carry out a 10-day ensemble prediction test and a randomized time and space scale sensitivity test in September and October 2018 (choose 1st, 7th, 13th, 19th, and 25th), and evaluate the test results. The main conclusions of the research work are as follows: By comparing the ensemble prediction results of the test using the SKEB scheme and the test without the SKEB scheme, the use of the SKEB scheme increases the large aerodynamic energy of the CMA-REPS regional model in the small and medium-scale region, and improves the CMA-REPS ensemble prediction system to the actual atmosphere to some extent. The simulation ability of kinetic energy spectrum; the introduction of SKEB scheme in regional ensemble prediction can significantly improve the spread of U and V in horizontal wind field of regional model. The SKEB program has improved the forecasting skills to a certain extent, such as reducing the CRPS scores of the horizontal wind fields U and V, reducing the outliers scores of the horizontal wind field, temperature, and 10 m wind speed; the introduction of the SKEB method can improve the light rain. The precipitation probability prediction skill score, but the improvement of the score did not pass the significance test, so it is considered that the SKEB method is difficult to improve the probability prediction technique of precipitation. Sensitivity tests based on the SKEB method for six time scales of random pattern (1h, 3h, 6h, 9h,12h, 15h of the time series τ) and five spatial correlation scale (the maximum cutoff wave number Lmax is selected as 80, 100, 120, 160, 200, 240) show that the ensemble prediction is sensitive to the six time scales of the stochastic model of the SKEB method. And the experiment whose time is set to 12h and the maximum truncation wave number is set to 240 show the best performance than the others. Therefore, we can draw the conclusion that the SKEB scheme can compensate for the small-scale kinetic energy dissipated at the truncated scale, and effectively improving forecasting skills.