Effective outlier detection for wind turbine power curves

Energy is an important factor in increasing the welfare level of societies. Fossil fuels, which are widely used in energy production, are decreasing day by day. This condition indicates that fossil fuel reserves will be depleted in the future, and leads to an increase in the use of renewable energy. Wind energy is a renewable energy source. It stands out in most geographical areas due to its medium- and long-term usage stability and low environmental impact. Wind turbine power curves are crucial for wind energy assessment, grid planning and troubleshooting. However, abnormal data found in the power curves needs to be cleaned in order to accurately model the power output. In this thesis, first of all, by using fuzzy c-means algorithm, both the raw power curves were partitioned into more than one cluster and the status of belonging to more than one cluster of existing data points were evaluated. Thus, abnormal data were analyzed more comprehensively. Afterwards, by computing the Mahalanobis distances for the clustered raw power curves, both abnormal data were evaluated as multivariate and available data points were colorized according to the level of abnormality. Thus, abnormal data were cleaned more effectively. In addition, the datasets, which were collected at 10-min intervals for 12 months and included seasonal effects, were used. As a result, the applied abnormal data detection approach provided reliable and efficient results for wind turbine power curves.