The improvement of wind farm"s output power prediction accuracy can greatly reduce the impact of wind power on the grid and improve the security and reliability of wind power integration. In this paper, the FKNN (Fast K-Nearest Neighbor algorithm) algorithm is proposed to improve the shortcomings of KNN (K-Nearest Neighbor algorithm) algorithm and is used for short-term wind power prediction. First, for each prediction sample, by using FKNN algorithm, which is based on the principle of similarity data, you can obtain the maximum priority queue of similar sample through traversing the set of training sample only one time. Then, gradually reduce the length of the priority queue to produce different size priority sub-queues of similar sample in which the majority class samples can be obtained and it"s average is used to predict the output power of prediction sample. Finally, the algorithm"s simplicity and practicality was fully proved through the prediction of a large amount historical data of a wind farm in Jilin Province.