Abstract: To improve the prediction accuracy of extreme precipitation events, such as high-intensity precipitation and rainstorms, and to meet the demand for high-resolution and high-frequency predictions, this study used the Himawari-8 infrared band brightness temperature data, China’s Automatic Weather Station precipitation data, and ERA5 reanalyzed physical quantity data. This study then identified the mesoscale convective system (MCS) by using the cloud top temperature data of the satellite; used the feature variables screened by the machine learning feature algorithm as model input data; established a classification model and regression model with the Stacking integrated learning algorithm to predict the precipitation in the next hour; tested and analyzed the prediction results with indicators such as the Threat Score, confusion matrix, and correlation coefficient. The model was used to predict the hourly precipitation of the 23.7 extreme heavy precipitation event in North China. The results showed that the cloud top temperature data obtained by the stationary meteorological satellite significantly affects the identification of the precipitation type and quantitative prediction of precipitation. Further, the classification-prediction algorithm performed well in rainfall nowcasting for the MCS, with a correlation coefficient and root-mean-squared error of 0.811 and 0.14 mm/h, respectively. These prediction results are better than those obtained by individual regression models. For the 23.7 extreme heavy precipitation in North China, the TS score reached 0.58, and the forecast of the spatial distribution of precipitation agreed well with the actual distribution. In general, the fusion model realized a stronger and more stable prediction performance, and its prediction effect for samples with large precipitation should be further improved.