Clustering Improves Wind Farm SCADA Data Filtering Accuracy

Wind farms generate vast amounts of Supervisory Control and Data Acquisition (SCADA) data, which often contains anomalies, transients, and specific operational modes. For many analytical applications, only data representing normal operation is useful, necessitating a robust filtering process. Traditionally, this filtering has been a manual, expert-driven task involving visual inspection, which is time-consuming and prone to human error. This research explores the potential of clustering algorithms to automate and enhance this critical data preparation step. The study evaluates the filtering accuracy of multiple clustering algorithms against manual expert filtering, using SCADA data from three turbines of an offshore wind farm. A key contribution is the introduction of novel evaluation metrics specifically designed for unlabeled data, ensuring robustness across various potential applications. The dataset used was particularly challenging, containing not only typical anomalies but also numerous non-evident outliers resulting from field tests. The findings indicate that cluster-based methods generally achieve higher accuracy than manual filtering, successfully identifying both obvious and subtle outliers. However, the performance, including the amount of data retained, varies significantly depending on the specific clustering model employed. The research emphasizes the importance of extending analysis beyond simple power curves, both in feature selection and in designing appropriate evaluation metrics. While expert involvement remains necessary, its extent can be significantly reduced compared to purely manual approaches.