Deep Reinforcement Learning Assisted Hybrid Five-Variable Modulation Scheme for DAB Converters to Reduce RMS Current and Expand ZVS Operation

Abstract

In order to enhance the conversion efficiency of dual-active-bridge (DAB) converters, a hybrid five-variable (HFV) modulation strategy with the aid of deep reinforcement learning (DRL) technique is proposed in this paper. Specifically, deep deterministic policy gradient (DDPG) algorithm is employed to train the agent with the purpose of the lower root mean square (RMS) current and the zero voltage switching (ZVS) operation of all power switches under various conditions. Moreover, due to the complexity of HFV modulation with five control degrees of freedom, frequency-domain analysis is used to directly derive the operating expressions of the converter. Based on these, the trained DDPG agent can output the optimal values of control variables, which can lower the RMS current and achieve the widest ZVS operation with reduced computation burden. A thorough comparison among related works in terms of the RMS current and ZVS range was presented. Finally, the effectiveness of the proposed method is verified by experiments on a built prototype.