Abstract: Support vector machine (SVM) is widely used in wind speed forecasting and the forecasted results are verified well. With the applications get more intensive, there exist two problems in SVM. One is it is too sensitive to noises and the other is it can not fully use the information included in the samples. In view of this, a fuzzy manifold-based support vector machine (FMSVM) is proposed in this paper to solve the above problems and further improve the generalization capability of SVM. FMSVM introduces the fuzzy techniques to decrease the influence of the noises. Meanwhile, FMSVM takes boundary data between classes, data distributions and manifold seriously. The comparative experiments show that FMSVM performs better than SVM on the wind datasets of a certain wind farm.