Energy-based machine learning framework for cyber security enhancement in smart grid networks

Digital communication and automated control of smart grid networks are increasingly dependent on digital networks, which are susceptible to cyber-attacks that can be extended to physical disturbances of the power system. The current intrusion detection techniques are mostly based on the patterns of cyber traffic and can hardly differentiate between malicious and legitimate changes of operating variations in the dynamic grid settings. This paper presents a machine learning model grounded on energy to combine cyber-layer measurements with energy-space dynamics in a single learning representation. The method suggested uses physical consistency constraints in the classification process, which is not the case of the conventional cyber-only detectors. A simulation based smart grid dataset of 10,000 samples, and four operating classes. The proposed framework has a final detection accuracy equal to 95.8% and an F1-score of 0.95, which is 2-5% points higher than the typical baseline methods. The ablation analysis also proves the fact that the energy-domain features and constraints imposed by physical plausibility can add up to a significant increase of performance. The research results suggest that informed learning that considers physical considerations is an effective and viable way of achieving credible cyber security enhancement in smart grid networks.