Abstract: Accurate and reliable wind power forecasting is of great significance for improving the wind power consumption rate of power systems. Reasonably combining numerical weather forecast data is an effective means to improve the accuracy of wind power forecasting.This paper proposes a method to improve the ultra-short-term prediction accuracy of long-short-term memory networks based on historical similarity. Firstly,the numerical weather forecast is used as the characteristic of the extreme learning machine to generate correction data;then,through the weighted gray correlation algorithm,the historical data that is similar to the feature of the point to be predicted is extracted,and the prediction results of the long-short-term memory network are evaluated and corrected. The calculation example uses the actual operation data of a wind farm in Colorado,USA for training and verification,and uses different correction methods for comparison. The results show that the optimization method based on historical similarity can improve the prediction effect of the short-term memory network and reduce the error fluctuation range,and verified the method in this article.