Anomaly detection in power dispatch data networks based on graph convolutional networks (GCN)

Abstract

A graph convolutional network (GCN)-based approach is developed for anomaly detection in power dispatch data networks, addressing the challenges of high-dimensional, non-Euclidean, and dynamic grid environments. The method begins with the construction of a graph that encodes both physical connectivity and operational interactions between dispatch terminals and substations. Feature engineering and Z-score normalization ensure the standardization of system state data and various sensors for robust model input. The multi-layer GCN architecture aggregates local and global neighborhood information in the network, which can identify distributed anomalies and isolated anomalies. Comprehensive experiments were conducted on a large-scale real dataset from a provincial power dispatching system. The results show that the model outperforms the competitive deep learning baselines and traditional statistical methods in terms of detection accuracy and interpretability. The framework helps with power grid reliability analysis, anomaly detection, and propagation tracking. The dependence on historical data normalization and the assumption of static network topology constitute limitations. Studies have shown that GCNs can be used for intelligent power dispatch and anomaly management.