Abstract: The wind farm power data contain a large number of abnormal data, which are difficult to reflect the real wind energy situation of the wind farm, affecting the accuracy of wind power forecasting, thereby affecting the decision-making of the power grid. In order to solve this problem, we analyzed the characteristics of abnormal data of wind farms, and divided them into accumulation type and scattered type. Based on the time series change point detection theory, whether the density ratio is a constant value was used as the judgment criterion for eliminating accumulation type abnormal data. The improved Kullback Leibler importance estimation program (IKLIEP) was used to eliminate the stacked abnormal data, and the quartile method was used to eliminate the scattered abnormal data. Finally, the method in this paper was applied to a 130.5 MW wind farm in western Mongolia. The experimental results show that the method in this paper can ne adopted to more effectively identify and eliminate abnormal data, the average recognition rate is increased by 6.19%, and the false recognition rate is reduced by 2.92%, which proves the verifying the effectiveness of the method.