Critical component analysis of cyber-physical power systems in cascading failures using graph convolutional networks: An energy-based approach

Power systems, with increasing integration into communication networks, have evolved to become complex and interdependent cyber-physical power systems that are highly vulnerable to cascading failures. These failures, due to their propagation through the cyber and physical networks, often lead to severe disruptions. We employ improved percolation theory to model cascading failures triggered by malware cyber-attacks. Addressing the vulnerability of CPPS requires a comprehensive analysis that spans both the structural and functional dimensions of CPPS. This paper introduces a novel framework for vulnerability assessment in CPPS using Graph Convolutional Networks (GCN). Our approach captures the topological complexities and dynamic characteristics of CPPS, incorporating the entropy of potential energy of power system as anew feature to predict and analyze failure propagation. Through Layer-wise Relevance Propagation (LRP), we subsequently quantify the influence of potential energy on system vulnerabilities. Critical components are identified by using LRP scores and an entropy weighting method (EWM). Simulation results based on the cyber-physical IEEE 39- bus and IEEE RTS-96 power systems as test cases, demonstrate the model's efficacy in identifying vulnerable nodes and branches and highlight the significant role of potential energy in cascading failures. This framework provides a comprehensive approach for real-time vulnerability assessments and resilience enhancement in CPPS.