In order to improve the accuracy of short-time severe rainfall forecasts, the LightGBM algorithm is applied to build the hourly precipitation forecasting model based on the precipitation observation data and CMA-GD model forecast products of Fujian Province from April to September in 2019 and 2020. Correction models are optimized by the feature processing, Bagging (bootstrap aggregating) and hyperparameter searching. Combined with AUC, AUPR and traditional classification indices, a series of experiments are designed to evaluate different modeling schemes and verify the applicability in short-time severe rainfall forecasting. The results show that all modeling schemes can improve the original numerical model forecast representing the high POD and FAR in varying degrees. Bagging can enhance the stability of model prediction, and the slightly unbalanced sub-training set contributes to the higher TS scores by reducing the FAR with the best TS score of validation set being about 17.5%. The largest contribution of feature variable to the classification information gain is K index, followed by 500 hPa dew point and time parameters. The ranking of experiment indices in good to bad order is random cross-validation, random hourly cross-validation and operational simulation test which indicates that the validity of correction models mainly result from the sample information at the same or adjacent moments. The dynamic fusion scheme of heterogeneous models based on logistic regression increases indices of static homogeneous models, which decreases at least 520 〖KG-*5〗000 false alarm samples with approximately 50% POD.