Unified Reinforcement-learning-assisted PI Gain Adaptation for CC-CV Control of an SPS-modulated Dual-active-bridge EV Fast Charger

Abstract

This paper proposes a unified reinforcement-learning-assisted PI controller for CC-CV charging of an SPS dual-active-bridge converter.A DDPG agent acts as a supervisory layer that adapts bounded current-and voltage-loop PI gains while preserving deterministic mode logic, anti-windup, phase saturation, and pulse generation.Five independently trained agents are evaluated as frozen policies against a screened fixed PI baseline.Under the nominal test condition, the five-seed RL mean preserves CC performance, with current IAE and ISE changing by 0.32 and 0.62 percent, while reducing CV-voltage IAE by 26.44 percent, ISE by 36.18 percent, and overshoot by 15.59 percent.Across eight matched randomized operating conditions, the five-seed RL mean gives lower CV-voltage IAE than the fixed PI in all eight cases; the condition-level mean paired difference (RL-PI) is -0.00724V•s with a 95% confidence interval of [-0.01293, -0.00156] V•s.All 40 randomized Seed-44 switching-model cases satisfy the imposed batteryside limits.Frozen-policy multirate evaluation shows nearly invariant nominal behaviour from 0.1 to 5 ms, and all 108 tested frozen-gain cycle-averaged local linear cases are stable.The results support bounded state-dependent PI gain adaptation, while HIL and experimental validation remain necessary.