Improving the accuracy of objective classification and spatio-temporal forecast of severe convective weather has always been a challenge in meteorological forecasting. This paper integrates mesoscale models with machine learning classification algorithms, achieving hourly forecasting of classified severe convection. The specific algorithm is as follows. First, the XGBoost classification algorithm and historical data over 10 years are used to establish a classification potential forecast model for severe convection. Secondly, by statistically analyzing the optimal spatial neighborhood radius and probability density distribution characteristics of CMA-SH9 model elements, and extracting element thresholds based on the combination of optimal scores, a spatial neighborhood graded element ingredient model is established. Finally, through joint discrimination, the machine learning classification method and the spatial neighborhood element “ingredient method” are integrated to establish hourly forecast models for thunderstorm gales and short-time heavy rainfall. The validation results show that this fusion algorithm significantly outperforms numerical model forecast results and national guidance products. For short-time heavy rainfall forecasts on an hourly basis over 24 h during 2021-2022, the average hit rate is 0.51 and the TS (threat score) is 0.15, which reflects the improvements of 82% and 36% respectively compared to the model. For thunderstorm gale forecasts on an hourly basis over 24 h, the average hit rate reaches 0.37 and the TS is 0.07, representing improvements of 68% and 133% respectively relative to the model (reflectivity factor ≥45 dBz). Thus, the forecast accuracy of thunderstorm gale is significantly improved.