Accurate short-term wind power forecasting modeling is very important to improve the economic and stable operation of new energy power systems. In view of the shortcomings of traditional forecasting methods in small sample learning, fine modeling, probability prediction and the influence of easily falling into local optimum, this paper firstly establishes the wind power forecasting model of Relevant Vector Machine (RVM) based on the theory of RVM. Then, aiming at the lack of local optimum mechanism and group memory function in Gravitational Search Algorithm (GSA), an APSO-GSA hybrid optimization algorithm is proposed to optimize the parameters of RVM model. Finally, a wind farm operation data in Northwest China is taken as an example to verify the effectiveness of the proposed method. The results show that the proposed method has higher modeling accuracy and faster convergence speed, and achieves accurate prediction of wind power at future time using a small number of samples and simple models. This work is supported by National Natural Science Foundation of China (No. 51667018 and No. 51367015).