Accurately predicting wind speed is of key importance to the operation of the power systems with wind power plants. In order to improve the accuracy and practicality of short-term wind speed prediction, aimed at the characteristics of random variation and tendency of wind speed sample sequences, a short-term wind speed prediction method based on improved mean generating function with fuzzy theory and back-stepping algorithm is proposed. This paper firstly improved the construction process of the mean generating function by fuzzy theory and back-stepping algorithm, and then combined it with the optimal subset regression algorithm to establish the short-term wind speed prediction model. The case analysis shows that, comparing the wind speeds forecasted by the proposed model with those forecasted by traditional mean generating function model and ARMA based model, the new prediction model can combine the advantages of the mean generating function, the fuzzy theory and the back-stepping algorithm, and can greatly improve the prediction accuracy, which has a broad prospect of engineering application.