Robust Control of Multi-Cellular DC-DC Converters Using Model-Free Deep Q-Network Reinforcement Learning

This paper proposes a model-free reinforcement learning-based control strategy for a three-stage multi-cellular DC-DC power converter. A Deep Q-Network (DQN) agent is developed to regulate the flying capacitor voltages and the load current through direct interaction with the converter, eliminating the need for an explicit mathematical model of the system. To enhance learning stability and control performance, the framework incorporates tailored state preprocessing and reward-shaping mechanisms that guide the agent toward balanced capacitor voltages and accurate current tracking. The proposed reinforcement learning controller is benchmarked against a conventional Sliding Mode Controller (SMC) under step reference conditions to evaluate its dynamic and steady-state performance. Simulation results demonstrate that the DQN-based controller achieves comparable transient response while providing improved steady-state accuracy and robustness across varying operating conditions. In addition, the learned control policy generates smoother switching actions, reducing oscillations and mitigating switching stress on the power devices. These results highlight the potential of model-free reinforcement learning as a robust, adaptive, and scalable alternative to conventional control strategies for advanced power electronic converters.