Abstract: Lightning is one of the natural disasters that have a major impact on human safety. Monitoring and early warning of lightning are important means to reduce harm. The present study uses the atmospheric electric field data and lightning location data in Huangpu District of Guangzhou to analyze relevant features that reflect the lightning trend, find early warning factors from them, and explore their correlation with lightning events within the electric field detection range. The experiment is based on the base classifier BP neural network, and establishes the ensemble model through Bagging and Adaboost respectively. The experiments show that in the early warning of lightning event with a time window of 30 minutes, the features extracted in this experiment reduce FPR and FNR by 16.83% and 15.19%respectively, compared with the features provided by the current research. Compared with the single BP neural network, the ensemble method reduces FPR by 11.46% and FNR by 4.73%, which shows that the method of feature extraction and ensemble learning can effectively improve the accuracy of the model in lightning prediction.