A Deep Learning-Based Cyberattack Detection Method for Transmission Line Differential Relays

Abstract

Cyberattacks on power systems have increased, posing serious threats to control systems and protective relays. Line differential relays (LDRs) are widely used to protect critical transmission lines due to their fast, selective, and sensitive operation. However, despite these advantages, LDRs remain vulnerable to cyberattacks as they rely on communications to exchange measurements, which can be compromised. This paper proposes a new deep learning-based cyberattack detection method to detect false-tripping and missed-tripping/fault-masking cyberattacks targeting LDRs. The proposed scheme relies solely on LDR's local measurements, enhancing its security compared to previous solutions, as local measurements are more difficult for hackers to manipulate. The proposed method is based on a deep learning neural network (DLNN), providing a robust model to protect LDRs from cyberthreats. The DLNN model is trained offline on a wide multi-state dataset that includes possible conditions of normal operation, internal faults, and nearby external faults. Additionally, the hyperparameters of the DLNN model are optimized using Bayesian optimization. To reduce complexity, a rule-based system is integrated to identify the type of potential cyberattack instead of incorporating all cyberattack scenarios into the DLNN training phase as done in previous studies. The performance of the proposed method is evaluated under various scenarios, including normal operation, faults, and cyberattacks. The results demonstrate the superiority and efficacy of the proposed scheme in detecting cyberattacks. The proposed scheme outperforms recent literature by achieving nearly 100% classification accuracy on the test dataset. Even under the worst-case scenario of measurement noise, the classification accuracy drops slightly to 99.3667%.