Machine Learning-Based Identification of Control Algorithm for Chalmers Test Wind Turbine

Abstract

Abstract Wind turbine control strategies are rarely publicly disclosed, limiting access to detailed information on internal algorithms. This hinders third-party analysis and understanding of turbine operation, which is essential for addressing the challenges and potential of wind energy integration in modern power systems. Therefore, datadriven methods are required to extract insights from available measurement data. The thesis presents a method to investigate the control behavior of a research wind turbine at Chalmers University of Technology using measured data. The approach is based on graphical analysis of torque versus rotor speed, employing measurements of DC current and voltage at the generator rectifier as well as rotor speed. To increase the reliability of the method, two data pre-processing strategies are implemented: a binning-based data density filter and a generator torque setpoint reconstruction using supervised machine learning. The effect of various factors on model accuracy is investigated, showing that including the input parameters current, voltage, pitch angle, and rotor speed yields the highest reconstruction precision. The method also includes features for detecting changes in control strategies and estimating controller-specific configuration parameters. Control change detection is performed using a one-sample hypothesis test, which exhibits a false positive rate of approximately 7.5% for the cases considered. Validation results show strong agreement between most identified and actual controller parameters of the Chalmers turbine under both pre-processing approaches indicating high method accuracy. For commercial turbines, reduced accuracy is expected due to simplifying assumptions, such as linearized optimal torque-speed relationships, and the unavailability of torque setpoints for supervised learning. Additional machine learning techniques were explored but did not lead to better performance. The proposed method offers a data-driven tool to analyze wind turbine control behavior and supports reverse engineering of control strategies where internal system information is unavailable.