Wind turbine power curves are greatly important for monitoring the turbine performance, forecasting the power generation and mitigating the system unreliability. However, the presence of abnormal values in the power curves affects the modelling of normal turbine behavior, conversely. This study explores the anomalies in the raw power curve data of a wind turbine and offers a robust outlier detection approach at three levels: (1) the k-means clustering based on Squared Euclidean and City-Block distance measures, (2) the silhouette computing to compare both clustering solutions and (3) the data filtering according to the Mahalanobis distance thresholds. As a result of all conducted levels, the proposed partitional clustering-based outlier detection approach has shown the good identification of abnormal data points and the refined power curve data is achieved, effectively.