Towards a more secure reconstruction-based anomaly detection model for power transformer differential protection

Introduction Cyberattacks against Power Transformer Differential Protection (PTDP) have the potential to cause significant disruption and widespread blackouts in power infrastructure. Recent literature has demonstrated how reconstruction-based anomaly detection models can play a critical role in enhancing the security of PTDP against such attacks. However, these models themselves are vulnerable to cyber threats. Adversarial sample generation is an example of a threat against reconstruction-based anomaly detection models. Methods To address this threat, we propose an approach for adversarial training of such models appropriate for PTDPs. We then review and compare the effect of adversarial training on the performance of four different model architectures. To demonstrate the efficacy of our proposed approach for improved security and performance in PTDP scenarios, the IEEE PSRC D6 benchmark test system is tested in an OPAL-RT environment. Results Simulation results show the effectiveness of the proposed method for improved detection of cyberattacks.