Can DC Converters Achieve Resilient Voltage Regulation with Reinforcement Learning?

This paper investigates the impact of reinforcement learning (RL)-driven control on the resilience of DC converters under severe high-impact fault scenarios. DC buck converters are critical components in DC microgrids, providing efficient voltage regulation for the DC load bus. Ensuring robust control of these converters, even in faulty and unpredictable situations, is essential for maintaining system stability and preventing potential damage. To address this, an RL-agent control method is developed using the Deep Q-Network (DQN) algorithm, designed to operate with a discrete action space while achieving performance comparable to continuous methods, thereby requiring less computational power and offering improved accuracy. To assess its effectiveness, the RL-based control method is benchmarked against a conventional Proportional-Integral-Derivative (PID) controller. Real-time implementations are conducted using Typhoon HIL real-time simulator and NI myRIO 1900 as the controller hardware, establishing a Controller Hardware-in-the-Loop (C-HIL) testbed. Two fault scenarios, severe voltage drop in the source and short-circuit across the load, are tested to evaluate the proposed controller's performance. Results indicate that the RL-based method restores circuit dynamics more swiftly after faults compared to conventional approaches, demonstrating robust and adaptive control capabilities in dynamic environments. These findings underscore the potential of reinforcement learning to enhance the resilience of modern power systems.