Action-Space-Oriented Reinforcement Learning Compensation for PI-TPS-Controlled Low-Voltage DAB Converters Under Input-Voltage and Load Variations

Abstract

This study investigates how the insertion point and parameterization of reinforcement-learning (RL) compensation affect a low-voltage dual-active-bridge (DAB) converter operated with triple-phase-shift (TPS) modulation. A proportional–integral (PI)-TPS controller is used as the common baseline. Two deep deterministic policy gradient (DDPG) compensation structures are compared: one directly corrects the three TPS variables, whereas the other corrects the high-level phase-shift command before the deterministic TPS mapping. The controllers are evaluated at the rated point and under fixed and stepwise input-voltage and load variations. Direct three-variable correction provides no consistent improvement in the investigated simulations, while phase-shift-level compensation provides a more favorable voltage-current tradeoff and preserves TPS coordination. At the rated point, the phase-shift-level scheme reduces the steady-state voltage error by 29.34%, and the full-transient peak inductor current from 69.82 to 36.84 A. Variable-condition results represent robustness after retraining rather than zero-shot generalization. Because the two RL formulations also differ in observations and rewards, the comparison concerns the complete compensation structures rather than action-space dimension alone.