The intermittent and fluctuating characteristics of wind energy pose a great challenge to the smooth operation of power grid, which causes grid enterprises to restrict wind power grid connection, resulting in curtailment behavior. Therefore, the real-time and effective prediction of wind power generation is critical for the development of wind power and the smooth operation of power grid. After analyzing several current prediction methods, a short-term wind power prediction model based on kernel principal component analysis-K-means clustering-gated recurrent unit (KPCA-K-means-GRU) is proposed. Multidimensional data can restore the real physical state better, but data with too high dimensions will cause dimension disaster. Therefore, a non-linear KPCA is used to reduce the data dimension while retaining the information of high dimension data. Then based on the idea of similar days for load prediction, unsupervised clustering of reduced dimension data by K-means is used to establish different prediction models to improve prediction accuracy. Finally, the GRU neural network parameters of different kinds of data are trained separately, and then classification prediction is carried out to obtain a more appropriate network model.