COMPARATIVE ANALYSIS OF SUPERVISED AND UNSUPERVISED MACHINE LEARNING FOR RELIABLE WIND TURBINE MONITORING AND SUSTAINABLE ENERGY GENERATION – ENERGY POLICY

Abstract

Condition monitoring plays a role in wind turbine maintenance. But supervised machine learning is constrained in the presence of unlabeled data in datasets related to Supervisory Control and Data Acquisition (SCADA). This paper suggests method to detect anomalies in wind turbines based on 8,760 observations per hour from the Texas Wind Turbine SCADA dataset. Power estimation is performed using a 300-tree model of Random Forest regression that uses wind speed, wind direction, air pressure, and air temperature. The power difference between the measured and expected powers is used to label the anomaly using a threshold of 4.1767 kW. The anomalous observations are identified as 700 and normal observations are 8,060. The optimized Random Forest and XGBoost models have a 93.06% and 93.24% accuracy, respectively, and their F1 scores are 41.54% and 43.51%, respectively. LOF finds 438 density anomalies, and McNemar's test shows that there is no significant difference between classifiers.