A Comparative Study of Smart Grid Security Based on Unsupervised Learning and Load Ranking

Due to the increasing number of risk factors, energy sector has been experiencing interruptions (attack) in the normal operation both externally and internally. Different methods are used for the identification and evaluation of vulnerabilities due to these interruption in the complex and critical infrastructures like the smart grid. Based on the objective of the attack, the performance and effectiveness of the learning-based approaches may vary when compared with other approaches to identify critical components of the smart grid. In this work, we adopted two target selection strategies (one is an unsupervised learning algorithm and the other is a load ranking based approach) for attack and measured system performances based on two evaluation metrics. We conducted the experiments on four different standard power system test cases and compared the performances of the aforementioned target selection strategies by two evaluation metrics. We used K-means clustering as the unsupervised learning method for the target selection of contingencies. To evaluate the system damage, we used generation loss and number of transmission line outages. For different attack orders, with two different attack objectives (evaluation metrics), experiments were conducted on W&W 6 bus system, IEEE 7 bus system, IEEE 8 bus system, and IEEE 300 bus system. We showed that, a clustering based attack performs better when the system is relatively large (highly dense in terms of connection to other buses) and the objective is to achieve a high number of transmission line outages. On the other hand, load ranking based attack outperforms clustering based attack when the attack objective is to achieve higher generation loss, regardless of the size of the system.