A Comparison Study of Feature Extraction and Data Fusion Techniques for Improving Cyber-Physical Situational Awareness

The power grid has historically been considered independently from the communication networks, however the physical system and the cyber system are becoming more intertwined as grid modernization initiatives push toward modern digital components. It is no longer sufficient to model the physical power system in isolation; the full cyber-physical system must be modeled for a complete system picture. Issues which were once purely cyber issues can now directly affect the physical system. This work investigates techniques for fusing cyber and physical data to analyze a scenario which includes a physical disturbance and a Denial-of-Service cyber attack which impedes control commands during the physical disturbance. Principal Component Analysis with Singular Value Decomposition, t-distributed stochastic neighbor embedding, and autoencoders are explored and compared for extracting features from cyber-only, physical-only, and cyber-physical data, qualitatively comparing the methods to provide cyber-physical situational awareness for the power system.