This study proposes a controller design methodology based on deep reinforcement learning for buck converter control, aiming to enhance control performance and reduce design effort. In the proposed approach, the agent considers multiple objectives through reward design. By conducting training without fixing the load value, the agent can acquire control policies adaptable to a wide range of load conditions rather than being limited to a specific one. Incorporating neural network nonlinearity into this design further enables the resolution of trade-offs among responsiveness metrics, thereby improving transient performance while achieving higher robustness against load variations, as demonstrated through simulation results. Additionally, the robustness of the proposed controller to modeling errors was validated by simulating scenarios with varied circuit parameters and evaluating the system responses. Moreover, the proposed method eliminates the need for trial-and-error processes in the current observer design and gain tuning. Consequently, it achieves both high control performance and reduced design burden—an outcome that has been challenging to realize with conventional methods. Therefore, this method is expected to contribute to the broader implementation of DC power distribution systems and microgrids.
Obayashi et al. (Sun,) studied this question.