Wind turbines abnormality detection through analysis of wind farm power curves

Abnormality detection and prediction is a critical technique to identify wind turbine failures at an early stage, thus avoiding catastrophes. In this study, we propose a new abnormality detection and prediction technique based on heterogeneous signals and information, such as output power signals and wind turbines downtime event information collected from the supervisory control and data acquisition (SCADA) system. First, discriminant statistical feature extraction is performed on the power signals in both the time-domain and frequency-domain. Then, a sideband expression is derived for normalized statistical data based on quartiles. In addition, a dissimilarity metric is defined to calculate the distances between downtime time intervals, and a higher dimension feature space is obtained. To reduce the dimension of the feature space, the Laplacian Eigenmaps (LE) nonlinear dimensionality reduction method is implemented. Afterwards, a Linear Mixture Self-organizing Maps (LMSOM) classifier is applied to differentiate abnormal types and a cumulative trend difference method is utilized to predict the faults in wind turbine. The method is validated and applied to data collected from a wind farm in north China. The results show that the proposed technique can effectively detect and predict wind turbine abnormalities.