Wind turbine performance analysis based on multivariate higher order moments and Bayesian classifiers

• Statistical filtering of a power curve based on k-means clustering.
• Wind farm analysis based on exploring higher order moments of their power curves.
• Quantitative analysis of higher order moments of a windfarm on different time scales.
• Equivalence between Hotellings T2 thresholds and measures in a Bayesian framework.

A data-driven model based on Bayesian classifiers and multivariate analysis of the power curve (wind speed vs. power) for monitoring wind farms' performance is presented. A new outlier detection criterion and various control bounds on the skewness and kurtosis of the data for cluster separation and classification of turbines' faulty and normal state of operation are introduced. Further continuous monitoring is addressed with Hotelling's T2 and Bayesian network approaches, and it is proven that under certain conditions, the outcomes of these two methods are equivalent. The Bayesian approach, however addresses the likelihood of classification, making supervised controls more flexible.