Anomaly detection in wind turbine SCADA data for power curve cleaning

Wind turbine power curve cleaning, by way of removing curtailment, stoppage, and other anomalies, is an essential step in making raw data useable for further analysis, such as determining performance, site characteristics, or improving forecasting models. Typically, comes SCADA (Supervisory Control Data Acquisition) data, so contains not only environmental performance but also the control action imposed on operator. Many different anomaly detection (AD) methods have been proposed to clean curves; however, few papers explored filtering explicit obvious anomalies from prior running AD. This paper actively explores this impact comparing performances 4 AD with/without filtering. These are: iForest, Local Outlier Factor, Gaussian Mixture Models, k-Nearest Neighbours. Each approach evaluated terms prediction error, removal rates, ability maintain underlying wind statistical characteristics. The results show effectiveness with every technique showing improvement compared its unfiltered counterpart. Furthermore, Models are shown provide favourable accuracy whilst maintaining variability, wide range methods, a user's choice may be depending their needs.