The invention discloses an ultra-short-term wind power prediction method considering the historical sample similarity, and the method comprises the following steps: analyzing the correlation between acurrent power value and a historical power value and the correlation between the current power value and a meteorological factor historical value, screening attributes with higher correlation, constructing historical samples, and reflecting the information of the power of a fan at the current moment. After dimension reduction of a historical sample matrix is conducted through a principal component analysis method, K-means clustering is carried out, and an appropriate clustering category K is selected according to a prediction effect, wherein K different clustering categories represent power generation conditions of different wind conditions; according to the category labels, historical numerical weather forecast information is adopted as input, the wind power value at the current moment is adopted as output, corresponding K support vector machine prediction models are established, and hyper-parameters such as the penalty coefficient and the kernel function bandwidth of the support vector machine are determined through a cuckoo search algorithm. According to the method, the problems that all external information cannot be reflected and overfitting are solved, the prediction precision can be effectively improved, and therefore the wind power absorption capacity is improved.