Deep Reinforcement Learning-Controlled GaN Dual-Active-Bridge Converter for Bidirectional V2GOn-BoardChargingin800VBEVPlatforms

Abstract

This paper presents the design, modulation, closed-loop control, and experimental validation of a 22kW, 500kHz Gallium Nitride(GaN) high-electron mobility transistor(HEMT)-based Dual-Active-Bridge (DAB) isolated DC-DC converter for full bidirectional Vehicle-to-Grid (V2G) and Grid-to-Vehicle (G2V) on-board charging (OBC) in 800V battery electric vehicle (BEV) platforms. A novel Extended-Phase-Shift (EPS) modulation strategy is augmented with a Twin Delayed Deep Deterministic Policy Gradient (TD3) deep reinforcement learning (DRL) controller that simultaneously optimizes efficiency, zero-voltage-switching (ZVS) range, RMS current stress, and grid total harmonic distortion (THD) across the complete bidirectional power envelope without hand-tuned PI regulators. Operating at a record 500kHz switching frequency with GaN Systems GS66516B devices (650V/60A) and a Vitroperm 500F nanocrystalline-core high frequency transformer, the converter achieves a power density of 7.6kW/L in a 210×140×28mm form factor. The TD3 agent is trained off line in a PLECS/Python co-simulation environment and deployed on an AMD Xilinx RFSoC ZU28DR SoC with 1.4µs FPGA-accelerated neural network inference latency. Experimental results confirm a peak round-trip G2V-to-V2G efficiency of 97.9%, grid-side current THD of 1.1% under V2G reactive power injection, ZVS maintained from 5% to 100% of rated power in both power directions, and 8ms dynamic power-set point settling time, with full compliance to ISO15118-20 Plug-and-Charge (PnC) V2G communication protocols. The proposed architecture deliversa 38% power density improvement and a 0.9% peak-efficiency gain over the best published SiC-based counterpart at equivalent power levels.