The historical operating data collected from wind farms，especially wind and power data，is significantly important for operation and management of wind farms and scheduling of a power system.However，wind curtailments are severe in practical operations of wind farms，causing large amounts of stacked abnormal data clusters distributed horizontally in a wind-power scatter diagram.This kind of data leads to large errors in an equivalent power curve and inaccurate wind power prediction，affecting wind farm management and power system scheduling.According to the characteristics of abnormal data，this paper presents a combined model for eliminating abnormal data based on the quartile method and cluster analysis.The quartile method is used twice to eliminate scattered abnormal data and cluster analysis is then used to eliminate the stacked abnormal data.Moreover，the problem brought about by “k”value selection in k-means clustering is solved by a novel “re-cluster”method.A case study shows that the model presented is efficient for eliminating abnormal data clusters and can often be used for both wind turbines and wind farms for its practical advantages.