In recent years, China's wind power industry has developed rapidly. However, wind power has instability, and the accuracy of ultra-short-term wind power prediction results directly affects the safe and efficient operation of the power grid. In order to further improve the accuracy of wind power ultra-short term prediction, a wind power prediction based on AM-LSTM model was proposed. This model combines long-term and short-term memory network (LSTM) with attention model (AM). In combination, the LSTM network can handle the nonlinear relationship between time series variables such as wind speed and wind direction and wind power, and the attention model can optimize the weight of the LSTM network to make the prediction result more accurate. Experiments using real wind farm historical data show that the proposed AM-LSTM prediction model can effectively utilize multivariate time series data for ultra-short-term prediction of wind farm power generation, which is more accurate than traditional BP neural networks and LSTM networks forecast effect.