Data-Driven Diagnosis of Wind Turbines via SCADA-Based Anomaly Detection Methods

The increasing demand for wind energy requires reliable predictive maintenance strategies to minimise premature failures caused by mechanical phenomena such as wear. Therefore, the early detection of failures is essential to improve the efficiency of predictive maintenance and ensure continuous electrical energy production. Over the last decade, data collected by the Supervisory Control and Data Acquisition (SCADA) system has become a common solution, especially due to its easily accessible data and cost-effective nature. This thesis focuses on developing various machine learning methods to establish an anomaly detection strategy, starting with the computation of a regression model, and concluding in an anomaly index that enables the assessment of the turbine’s health status. Specifically, this approach combines a physical analysis based on the Betz model with a machine learning technique, namely Support Vector Regression, to distinguish normal behaviour from potential faults. Additionally, the advantages and main limitations of a univariate analysis are investigated, and a comparison with a multivariate analysis is also performed. The workflow performance is evaluated through the application of several performance metrics typically used in this field to assess the model’s ability to detect anomalies under normal operating conditions. This standardised approach facilitates comparison with other techniques as defined in related research. Finally, a classification method involving multiclass analysis is conducted, with the computation of an additional performance metric: class error. An extract of this thesis was also presented during the SURVISHNO 2025 conference in Paris, under the title Anomaly Detection in Wind Turbines under Operational Variability via SCADA and Residual Analysis.