One approach to improve the accuracy and reliability of ultra-short term wind speed prediction is to thoroughly exploit the characteristics and laws of wind speed correlation from historical observations. The reference vector of k-nearest neighbor prediction based on spatial correlation is formed by the combination of latest local wind speed historical observations and upstream wind speed observations adjusted by its optimal lag time. The correlation coefficient is taken as the concrete evaluation index

and k most similar neighbours of the reference vector are optimally selected from the wind speed historical observations. Seven regression models are adopted for the future local wind speed prediction. The simulative results of wind speed prediction of Huibertgat

Holland in winter show that the optimal number of k-nearest neighbours is about 100 and the optimal year number of historical data is 10 a by the prediction of three optimal models

i.e. linear regression

partial least squares regression and least squares support vector machine regression

and the proposed method can effectively use the similarity of historical data for reliable ultrashort term wind speed prediction.