To insight the performance of numerical weather prediction model on refined heavy rainfall, a neighborhood verification method named Fractions Skill Score (FSS) is introduced to examine the capability of six operational models on 3h-accumulated precipitation during the warm season in 2019 in Zhejiang Province, with particular focus on short-term torrential rain and the impacts of different weather backgrounds. The results show that: (1) Based on the cumulative climatological probability of station 3h-accumulated precipitation, the FSS method is improved by determining the skill score thresholds for 5 grades precipitation with thresholds of 0.1, 3.0, 10.0, 20.0 and 50.0 mm/3h as 0.583, 0.522, 0.506, 0.502, and 0.500, respectively, and applied in the assessment of their prediction skill scales in the long time series. (2) The regional models outperform the global models in heavy rain forecasts, only the mean scores of Shanghai Regional Numerical Weather Prediction System (CMA-SH9) and Zhejiang Regional Numerical Weather Prediction System (CMA-ZJ3 and CMA-ZJ9) achieve the forecast skills for torrential rain, and their corresponding skill scale is 159, 159, and 183 km. There are about 60% of forecasts in these models achieve the prediction skills, and the cumulative frequency of the skill scale is increased by nearly 50% from 3km to 183km. This scale-selective evaluation result can provide a reference for the application of model products at different scales. (3) The differences in the models performance of precipitation forecasts under various weather backgrounds are obvious. The best model for predicting torrential rain under the background of tropical cyclones, Mei-Yu fronts and weak-synoptic forcing is the Global Forecast Model from European Centre for Medium-Range Weather Forecasts (EC-GFS), the CMA-SH9 and the CMA-ZJ3, with skill scales as 27, 99, and 135 km. Therefore, the application of model products should be treated differently in terms of weather background type.