Empirical Comparison of Machine Learning Algorithms for Mitigating Power Systems Intrusion Attacks

The normal and stable operation of the modern power systems rely on accurate situational awareness and visibility as recent researches and experiences have shown that the cyber-physical infrastructures are highly vulnerable to cyberattacks and intrusions. Attackers can design various intrusive injections to disrupt the operation thereby triggering failures, loss of synchronism, economic losses and sometimes injuries to employees. Hence, there have continuously been crucial need for timely, accurate identification and detection of these intrusions. Several traditional intrusion detection systems proposed in the literature have proven inefficient as they are computationally incompetent for the complex nature of the modern power systems. An alternative has been identified in form of machine learning techniques. This paper presents an empirical comparison of five prominent machine learning algorithms: K-nearest neighbors, Decision Tree, Naive Bayes, Random Forest and AdaBoost for predicting intrusion attacks into power systems network. The idea is to present the best possible classifier for the analyzed test systems and also to show that each of the developed algorithms can perform exceptionally well within some context. The developed algorithms were evaluated using a simulated voltage dataset generated from a load flow analysis of a 24-bus power systems case study. Experimental analysis and results obtained showed the feasibility of applying machine learning techniques in successfully predicting and detecting intrusions into power systems network.