Multimodal Learning in Cyber-Physical System: A Deep Dive with WSCC 9-Bus System

Multimodal learning has been utilized for numerous applications in cyber-physical power systems for load forecasting, voltage control and power quality monitoring. Although this learning technique has attained prominence in recent years, the use of this technique for predicting physical anomaly for cyber disruption is rather limited. This study uses the knowledge of t-distributed stochastic neighbor embedding (t-SNE) as a method widely used in multimodal learning for predicting the state of operation of a cyber-physical system. Naturally the aim of a power system is to keep the system operating which is termed as the normal or stable operating state of the power system. In the event of a possible contingency, the cyber-physical power system should foresee such event and can take preventive measures considering the comprehensive knowledge pertinent to this study. The Western System Coordinating Council (WSCC) 9-bus power system is taken as the test case for this study. In the event of loss of a generator, the system is brought into stable operation by load shedding. In the meantime, the system also underwent a Denial of Service (DOS) attack which denied the load shedding command. As a result, there was physical impact in the system which caused the system to become unstable. This is mitigated by issuing firewalls, which ensures the load shedding command will be initiated. Fusing cyber-physical data from multimodal relays and sensors located at strategic locations of the power system, the analysis is performed. The fused data is used to give an accurate estimation of the operating state that is visualized with t-SNE.