Due to the complexity of lightning formation and development process, the lightning proximity warning based on single data is difficult to meet the requirements in terms of prediction accuracy. In this paper, we conduct a lightning activity warning study based on lightning localization data and meteorological data. The combination of ERA5 dataset and lightning localization dataset is used as the model training and testing data, and the format unification of 1 km spatial resolution and 10 min temporal resolution is achieved with the help of cubic spline interpolation. On this basis, this paper compares the good and bad prediction performance of decision tree, support vector machine and bp neural network models for the target dataset, respectively, in which the support vector machine model performs better in terms of probability of detection (POD), while the bp neural network performs better in terms of false alarm rate (FAR).