A Deep Reinforcement Learning Framework for 3L-NPC-DAB Converters With Multiple-Degree-of-Freedom Phase-Shift Control

The three-level neutral-point-clamped dual-active-bridge (3L-NPC-DAB) converter offers several advantages over the conventional two-level DAB converter, including enhanced control flexibility and the ability to operate with lower voltage-rated power switches. However, with the increase in the number of power devices and multi-objective control requirements, traditional phase-shift control cannot meet the requirements. In order to further improve the conversion efficiency, this study introduces a deep reinforcement learning (DRL) optimization scheme based on the five control degrees of freedom (5-DoF) technique. The proposed scheme utilizes the deep deterministic policy gradient (DDPG) algorithm to minimize power losses and determine optimal control solutions. Through training, the DDPG agent acts as a predictor, enabling effective control decisions for maximizing conversion efficiency across various operating conditions. The effectiveness of the proposed method is validated through experimental results obtained from a laboratory proto-type.