The quality of wind power data affects wind power prediction and WTG output modeling. Due to mechanical failures, sensor errors and other reasons, the data collected contain abundant outliers, and it is particularly important to identify the outliers. In this paper, we propose an adaptive identification method, which includes the physical characteristics method to deal with outliers stacked up at the bottom along the direction of wind speed, the EM clustering (Expectation-maximization algorithm, EM) algorithm based on the profile coefficients to identify discrete outliers, and the Doseresp function based on the identification of outliers to identify discrete outliers. On the basis of identification, the Doseresp function fitting curve is used as a benchmark to reject the outliers, and finally the effectiveness of the proposed method is verified by taking a cluster wind farm in Gansu as an example.