Method for Cleaning Abnormal Data of Wind Turbine Power Curve Based on Density Clustering and Boundary Extraction

• Published in: IEEE transactions on sustainable energy 2022-04, Vol.13 (2), p.1147-1159
• Main Authors: Luo, Zhihong, Fang, Chengyue, Liu, Changliang, Liu, Shuai
• Format: Article
• Language: English
• Subjects: Algorithms; Boundary extraction; Cleaning; Clustering; Clustering algorithms; Data analysis; Data mining; Density; density clustering; Indexes; outlier cleaning; Outliers (statistics); Performance evaluation; raw SCADA data; Regularization; Turbines; Wind farms; Wind power; wind power curve; Wind power generation; Wind speed; wind turbine; Wind turbines
• ISSN: 1949-3029; 1949-3037
• DOI: 10.1109/TSTE.2021.3138757

This paper creatively proposes a complete set of procedures to identify and eliminate outliers of wind-power data based on the framework of classification processing. Outliers are divided into three types. According to the characteristics of each type of outlier, separate suitable algorithms are proposed. Through a comprehensive and in-depth analysis of the bottom stacked points, several operating modes of wind turbines at ultra-low wind speeds are discovered, and an intuitive rules method is innovatively proposed based on mechanism analysis. With the intuitive rules method, normal points and type 1 outliers are separated accurately in a reasonable way, and then type 1 outliers are cleaned up completely. For type 3 outliers, an improved density clustering method is proposed that makes full use of the density difference between normal points and type 3 outliers to achieve better cleaning performance. For type 2 outliers, a combined method including boundary extraction and boundary regularization is proposed. Type 2 outliers are completely removed as abnormal bulges on the boundary. Twelve datasets from three wind farms were employed to verify the effectiveness, superiority, and strong generalization of the proposed method. The power curve models based on smoothing spline method are used to visually display the effect of outlier cleaning. The identification accuracy (P), identification rate (R) and overall accuracy (F1) are introduced to quantitatively evaluate the cleaning performance.