In order to adjust the grid dispatching plan in real time and improve the grid"s ability to absorb wind power , a new ultra-short term wind speed prediction method is put forward based on dynamic time warping (DTW)-fast correlation-based filter(FCBF) and long-short term memory neural network(LSTM). Firstly, similarity of historical wind speed data is analyzed by using DTW method, and high similarity data to the predicted day are selected. Secondly, the FCBF method is used to select the optimal input feature set. Finally, the LSTM model is applied to predict the ultra-short term wind speed. The measured data of wind farm are forecasted, and the effectiveness and superiority of the proposed method is verified by comparing the prediction error with existing methods.