Cyber-Attack Detection in Smart Grids: A Comparative Analysis of Supervised and Semi-Supervised Methods

The advancement of smart grids has addressed many challenges of traditional power grids, yet it has also introduced new vulnerabilities to cyber-attacks that can disrupt power, leading to severe socio-economic impacts like blackouts and grid disturbance. While numerous supervised machine learning methods have been proposed to detect cyber-attacks in smart grids, they require a large dataset of normal and attack instances for training. However, gathering sufficient samples of diverse attack scenarios, especially zero-day attacks, is challenging. In this paper, we develop a semi-supervised model to detect grid attacks using phasor measurement unit (PMU) data from a power system dataset. Principal component analysis (PCA) is applied to select optimal components, and the model is trained using high instances of normal event data, thus enabling it to identify new, unknown attack patterns. We also developed a supervised model for comparison, evaluating both using key metrics. Results demonstrate that the semi-supervised model is more effective in detecting attack events (with 91.2% precision and 90% accuracy) than the supervised approach (90.7% precision and 91.8% accuracy).