The quality of wind power sample data and the structure of a wind power prediction model directly affect the accuracy of wind power prediction. A high-precision ultra-short-term wind power prediction method combining cross Local Outlier Factor (LOF) and an attention mechanism is proposed. Outlier detection is of wind power data at minute level by a cross LOF algorithm. This effectively improves the quality of sample data to avoid the loss of data characteristics caused by a fixed length vector in the encoding and decoding process of the Long Short Term Memory (LSTM) algorithm. An attention mechanism is added, so that the characteristics of historical data can be used more effectively to improve the accuracy of wind power prediction. Finally, the feasibility and accuracy of the method proposed are verified by the experimental analysis of real wind field measured data. This work is supported by National Natural Science Foundation of China (No. 51677117) and Key Research and Development Program of Shandong Province (No. 2019JZZY020804).