Graph Neural Networks for Anomaly Detection in Wind Turbines

The development of methods for anomaly detection in renewable energy systems is challenged by the complex spatio-temporal correlations among the measured signals, the variability of the operational and environmental conditions, and the presence of control systems that may hide anomalies in the signals patterns. To address these challenges, in this work we use Graph Neural Networks in identifying variations in the relationships among signals. Specifically, we propose a method which employs graph attention networks to dynamically consider interdependencies among signals, and gated recurrent units to capture temporal dependencies and system dynamics. The proposed method is compared to other state-of-the-art methods on a synthetic case study based on simulated wind turbine data. The obtained results show that the proposed method outperforms the comparison methods by more than 10% in terms of accuracy and reduces the detection delay by more than 50%.

Keywords: Renewable energy systems, Wind turbine, Anomaly detection, Graph neural networks.