The random forest algorithm is currently one of the more widely used machine learning methods, featuring high prediction accuracy, stable training results and generalization ability, and has obvious advantages in solving the problem of multiclassification. This paper applies the random forest algorithm to the prediction and classification of severe convective weather, which is divided into four categories: shorttime heavy rainfall, thunderstorm gale, hail and no severe convection. Then, based on the data of convection index and physics calculated from the NCEP data of 2005-2016, the training, 0-12 h forecasting and testing of classified severe convection are carried out. The results show that the whole misjudgment rate is 21.9% that is calculated out of the independent data of 2015-2016. It has almost no omission in 85 examples of severe convective weather and the model is especially suitable for larger range of severe convective weather. The physical meaning of the factors used in random forest algorithm is relatively clear, and basically consistent with the subjective forecasting experience. It can be used in daily forecasting operation.