The Security of Reinforcement Learning Systems in Electric Grid Domain

Abstract

Over the past few years, reinforcement learning (RL) has seen wide-scale applicability and possible deployments within power grids. Some of the applications of RL include control, cybersecurity, privacy, optimization, and energy management. The main advantage that RL offers within these areas is its ability to deal with unknown uncertainties via dynamic, experiential learning. This means that when RL encounters situations that it has not been trained on before, it stores them as experiences in memory and trains itself to deal with them in the future. In this chapter, we explore the typical working principles of RL and demonstrate its use as a controlling entity via a case study of secondary control within autonomous, multiagent alternating current microgrids. We also present simulation results to provide a visual depiction of the advantages achieved as a consequence of RL deployment for control. Additionally, we present an extensive discussion on how RL-enabled power grids can be manipulated with adversarial attack vectors and how they can compromise nominal operations.