Abstract: An ultra-short-term wind power forecasting method combining multiple clustering algorithm and hierarchical clustering algorithm is proposed. To deal with the dynamic condition of training samples and identify the samples that are similar to the characteristics of the period to be predicted, the historical power series and historical meteorological series are clustered separately.The clustering index of the power series consists of Euclidean distance and covariance, and the layer-by-layer method is used for meteorological series clustering. Two clustering results are combined into multiple sample subsets. Multiple neural network forecasting models based on particle swarm optimization and back propagation（PSO-BP） are established by using the method of classification modeling and feature matching. And the model with the most similar characteristics to the predicted period are used.The proposed forecasting approach has been applied in actual wind generation data tracking in Qinghai province of China. The simulation results show that it can improve the forecasting accuracy of ultra-short-term wind power.