Effectiveness of learning algorithms with attack and defense mechanisms for power systems

Abstract

Machine learning and artificial intelligence have created a myriad of opportunities for power systems. For example, learning techniques have been widely used for forecasting, data-driven frequency and voltage control, clustering of demand profiles, state estimation, theft detection, etc. Despite their satisfactory performances, these methods can face challenges associated with adversarial attacks that may result in system failure. In this paper, we systematically compare the effects of a set of attacks using different learning algorithms for a set of power systems problems (classification, forecasting, and control). We first show that most of the machine learning models proposed for power system applications are vulnerable to adversarial attacks. Then we adopt a series of common defense algorithms to improve performance under the different attacks. We identify GAN as the most effective defense algorithm across all power system applications tested, among all attacks and attack levels. It is of interest to notice that for the three power system problems different application datasets can be much more vulnerable. As the first systematic study of attack and defense mechanisms for a variety of power systems applications we hope to motivate the power systems and machine learning communities to continue studying how to increase the robustness of learning models against attacks. • We study the effects of attacks using different learning algorithms for power systems. • We test which defense algorithms are effective based on the attack and case study. • GAN is the most effective defense algorithm for all power system applications tested. • The power systems applications we study are classification, forecasting, and control. • Attack vulnerability varies widely based on the dataset studied in the control case.