Abstract: In order to solve the problems of sample scarcity and low forecasting accuracy of a single model in extreme weather, a short-term wind power forecasting method that integrates multiple models in extreme weather is proposed. Firstly, the original data of extreme events are extracted, and the feature redundancy and complexity are reduced by maximal relevance minimal redundancy (mRMR) feature selection strategy considering Granger causality. Secondly, to solve the problem of scarcity of extreme weather data, we use the time-series generative adversarial network (TimeGAN) algorithm to capture the dynamic characteristics of the data. Finally, the differences and advantages of each single model are taken into consideration, and an integrated model is constructed, in which the convolutional neural network, long- and short-term memory network, extreme gradient lifting tree, K-nearest neighbor algorithm, support vector machine are taken as the learning tools and lightweight gradient elevator is taken as the meta-learning tool, so as to predict the wind power in the next three days. Experimental results show that the proposed method can be adopted to effectively improve the short-term wind power prediction accuracy under extreme weather conditions, and the normalized mean absolute error and root mean square error are improved by 2.48% and 3.47%, respectively, compared with other prediction models.