In this study, based on the methods for object-based vertification and neighborhood, the hourly precipitation data from the CMADaas （China Meteorological Administration Data as a server）, the forecast performances of three numerical models during the main flood seasons from 2019 to 2020 in Liaoning Province are investigated. The three models are the Shanghai Numerical Prediction Model of China Meteorological Administration (CMA-SH9), the Mesoscale Weather Numerical Prediction System of China Meteorological Administration (CMA-MESO) and the Rapid-Refresh Multi-Scale Analysis and Prediction System-Northeast China Model of China Meteorological Administration (CMA-DB). The results show that even for the kilometer-scale or near-kilometer-scale models, there are still obvious deviations in the forecast of the heavy rainfall area (12 h cumulative rainfall ≥ 50 mm) within the 36 h lead time. The ratio of overlapping area between the forecasted and observed heavy rainfall areas is generally less than 10% of the total area, and in individual cases the value is close to 20%, such as the forecast of CMA-MESO on cyclone-type precipitation processes. The deviation of rainfall area results in a high missing alarm rate (MAR) (generally around 75%, and the MAR of the CMA-MESO is 10%-20% lower than that). The MAR of heavy precipitation forecast for the rear of high-pressure type precipitation exceeds 80%. Besides, the deviation of heavy rainfall area also results in a higher false alarm rate (FAR). For the forecast of short-term heavy rainfall (1 h rainfall ≥ 20 mm), by analyzing the mean values of the statistical indexes within 12 h forecast lead time, we find that the average percentage of detection is below 10%, with the maximum value being only 9.2%. The average FAR is 58.7%. Among the three types of rainfall processes, the model has the poorest performance in the forecast of short-term heavy rainfall in typhoon-type rainfall processes.