Vulnerability Evaluation of Multi-Agent Reinforcement Learning-Based Distributed Secondary Voltage Control in Microgrids
Conference
Taher, MA, Sarwat, A. (2026). Vulnerability Evaluation of Multi-Agent Reinforcement Learning-Based Distributed Secondary Voltage Control in Microgrids
. 10.1109/GreenTech68285.2026.11471556
Taher, MA, Sarwat, A. (2026). Vulnerability Evaluation of Multi-Agent Reinforcement Learning-Based Distributed Secondary Voltage Control in Microgrids
. 10.1109/GreenTech68285.2026.11471556
Multi-agent reinforcement learning (MARL) has emerged as a promising approach for distributed secondary voltage control in inverter-dominated microgrids due to its adaptive and model-free characteristics. However, the reliance of cooperative learning on shared measurements, communication, and reward feedback introduces cyber-physical vulnerabilities that are not yet well understood. This paper investigates the behavior and vulnerability of a cooperative multi-agent actor-critic-based secondary voltage control framework under cyberattack conditions on an IEEE 14-bus islanded microgrid with six inverter-interfaced distributed energy resources. Two adversarial scenarios are examined: state measurement false data injection and actuator-level voltage reference manipulation. Simulation results show that both attack types significantly disrupt agent coordination and degrade voltage regulation performance, leading to instability despite satisfactory operation under normal conditions. The study highlights fundamental limitations of conventional cooperative reinforcement learning controllers in adversarial microgrid environments and motivates the development of attack-resilient learning-based control strategies.
