Fleet-Wide Knowledge-Discovery-Based Methods for Wind Turbine Performance Monitoring: A Test Case Discussion

Wind turbine performance monitoring is essential to maximize wind energy conversion, but it is a task far less trivial than expected. In this work, a novel fleet-wide multi-step methodology is formulated for interpreting and classifying the behavior of operating wind turbines. At first, the monthly power curve profiles of each wind turbine in a fleet are elaborated to highlight anomalies. The wind farm is then categorized into two groups: potentially anomalous and presumably not anomalous wind turbines. Subsequently, fleet-wide methods are applied. Two Machine Learning regression models predict the wind speed (for each available sensor) and the power produced by each potentially anomalous wind generator, relative to the quantities measured at the surrounding turbines. By analyzing the residuals between the predicted and the actual values, relative performance changes and-or sensor loss of calibration are identified. One year of data from ten industrial multi-MW wind turbines are analyzed. One wind turbine results being affected by anemometer defective sensors and a slight performance improvement is highlighted in correspondence of the sensor recalibration, differently to a naïve scrutiny of the power curve profiles. Thus, the collected evidences highlight the importance of advanced data analysis methods for a correct interpretation of wind turbine operation.