Graph-enhanced bidirectional GRU for anomaly detection in power generation environments

Abstract This paper presents a graph-based anomaly detection model tailored for power generation systems using multivariate time-series data. The model integrated a bidirectional gated recurrent unit with graph neural networks and graph attention networks to effectively capture temporal dynamics and spatial relationships among sensors. The bidirectional gated recurrent unit enhanced prediction accuracy by considering historical and future information, while the graph neural network and graph attention network modeled component interdependencies, assigning attention weights to highlight critical connections. This approach addresses challenges inherent in traditional anomaly detection, such as data imbalance and the difficulty of identifying rare anomalies, by employing a prediction-based strategy instead of classification. The performance of the proposed model was evaluated using data from an industrial power generation control system, achieving detection of 44 out of 45 simulated abnormal events. The experimental results indicated that the bidirectional model surpassed unidirectional models, demonstrating the advantages of incorporating both spatial and temporal dependencies. Furthermore, the attention mechanisms improved the model interpretability, offering valuable insights for fault diagnosis. This study provides an efficient and interpretable solution for enhancing anomaly detection in power generation systems, contributing to improved safety and reliability.