Abstract
Conventional extended phase-shift (EPS) control methods for dual-active-bridge (DAB) converters are commonly designed under fixed voltage-conversion conditions, which limits their ability to exploit voltage-ratio-dependent efficiency improvement. In addition, accurate closed-loop power regulation remains challenging due to nonlinear switching dynamics and the high-frequency alternating inductor current. This paper proposes an optimal EPS control framework with residual reinforcement learning (RL) adjustment, in which offline optimization jointly determines the EPS parameters and voltage conversion ratio, while residual RL provides online corrections to enhance output-power regulation. The joint optimization is formulated to improve efficiency by minimizing backflow power and converter losses subject to zero-voltage switching (ZVS) constraints. A Twin Delayed Deep Deterministic Policy Gradient (TD3)-based residual RL controller is applied on top of the precomputed optimization-based baseline to compensate for tracking errors and operating uncertainties. To improve operational safety, the RL action space is constrained by hard boundaries derived from the steady-stateoptimization results. Experiments on a DSP-controlled DAB hardware platform demonstrate that the proposed method improves tracking performance compared with optimal baseline EPS, PI, and MPC methods, maintains stable operation under parameter variations, maintains ZVS operation, and achieves higher efficiency improvement over related methods without joint optimization.

This publication has 22 references indexed in Scilit: