Comparative Analysis of Model-Free Deep Reinforcement Learning Controllers for Reconfigurable Battery Systems Output Voltage Regulation

Maintaining the battery pack voltage at the reference value is crucial for the satisfactory operation of reconfigurable battery systems. This paper proposes a comparative study of reinforcement learning with various agents in reconfigurable battery system applications, focusing on four agents: deep deterministic policy gradient (DDPG), twin-DDPG (TD3), trust region policy optimization (TRPO), and proximal policy optimization (PPO). We evaluate and compare the results obtained by the agents with the conventional control method (PI). The simulation results in MATLAB confirm the superior performance of the studied methods, among which DDPG can even outperform other types of reinforcement agent.