Eigenvector centrality-enhanced graph network for attack detection in power distribution systems

Robust attack detection is critical for ensuring the reliability and security of power systems, which are increasingly vulnerable to sophisticated cyber–physical disruptions. Traditional detection methods often struggle to address the complexity and dynamic nature of modern power networks. This study expands on previous research that demonstrated the effectiveness of graph neural networks (GNNs) compared to other machine learning (ML) models in power systems security. We propose an eigenvector centrality-enhanced graph convolutional network (EVC-GCN) for attack detection, which improves the neighborhood aggregation process by assigning larger weights based on the global importance of nodes. This enhancement enables the GCN to better capture the topological significance of nodes in the network, leading to more robust and reliable attack detection. The method is evaluated across various attack scenarios on twelve power distribution networks of various sizes, using synthetic models for advanced, realistic testing of distribution systems (SMART-DS). The results demonstrate its superior accuracy and robustness in attack detection compared to conventional GCNs and other widely used ML models. • EVC-GCN emphasizes node importance in the neighborhood aggregation process. • EVC-GCN outperforms GCN and ML models when tested on 12 power networks. • EVC-GCN improves detection performance at the expense of slight increase in training time. • EVC-GCN scales well as network size grows, maintaining strong detection performance.