Condition Monitoring with Machine Learning: A Data-Driven Framework for Quantifying Wind Turbine Energy Loss

TL;DR

This paper presents a scalable machine learning framework for wind turbine condition monitoring that detects anomalies, isolates normal behavior, and estimates energy losses to improve maintenance and reduce economic impacts.

Contribution

It introduces a novel data-driven framework combining preprocessing, anomaly detection, and predictive modeling for effective wind turbine condition monitoring.

Findings

Significant data reduction retaining 31% of original SCADA data

24 turbines showed performance decline, 7 improved, 4 unchanged

Models like Random Forest and XGBoost effectively detect performance declines

Abstract

Wind energy significantly contributes to the global shift towards renewable energy, yet operational challenges, such as Leading-Edge Erosion on wind turbine blades, notably reduce energy output. This study introduces an advanced, scalable machine learning framework for condition monitoring of wind turbines, specifically targeting improved detection of anomalies using Supervisory Control and Data Acquisition data. The framework effectively isolates normal turbine behavior through rigorous preprocessing, incorporating domain-specific rules and anomaly detection filters, including Gaussian Mixture Models and a predictive power score. The data cleaning and feature selection process enables identification of deviations indicative of performance degradation, facilitating estimates of annual energy production losses. The data preprocessing methods resulted in significant data reduction, retaining on average 31% of the original SCADA data per wind farm. Notably, 24 out of 35 turbines exhibited clear performance declines. At the same time, seven improved, and four showed no significant changes when employing the power curve feature set, which consisted of wind speed and ambient temperature. Models such as Random Forest, XGBoost, and KNN consistently captured subtle but persistent declines in turbine performance. The developed framework provides a novel approach to existing condition monitoring methodologies by isolating normal operational data and estimating annual energy loss, which can be a key part in reducing maintenance expenditures and mitigating economic impacts from turbine downtime.