Deep-learning-based data-manipulation attack resilient supervisory backup protection of transmission lines

Cyber-attacks on smart-grid systems have become increasingly more complicated, and there is a need for taking detection and mitigation measures to combat their adverse effects on the smart-grid infrastructure. Wide area measurement system (WAMS) infrastructure comprising of phasor measurement units (PMUs) has recently shown remarkable progress in solving complex power system problems and avoiding blackouts. However, WAMS is vulnerable to cyber-attacks. This paper presents a novel cyber-attack resilient WAMS framework incorporating both attack detection and mitigation modules that ensure the resiliency of PMU data-based supervisory protection applications. It includes deep learning-based Long Short Term Memory (LSTM) model for real-time detection of anomalies in time-series PMU measurements and isolating the compromised PMUs followed by Generative Adversarial Imputation Nets (GAIN) for the reconstruction of the compromised PMU's data. The corrected PDC data-stream is then forwarded to the decision-making end application, making it resilient against attacks. A Random Forrest classifier is used in the end application to distinguish fault events from other disturbances and supervise the third zone of distance relay for backup protection of transmission lines. The efficacy of the proposed framework for different attack scenarios has been verified on the WSCC 9-Bus System modeled on a developed real-time digital simulator (RTDS)-based integrated cyber-physical WAMS testbed. Experimental analysis shows that the proposed model successfully detects and mitigates attacks' adverse effects on the end application.