A Model-Free Deep Reinforcement Learning Framework for Efficiency Optimization of Three-Phase Dual Active Bridge Converters

The three-phase dual active bridge converter is widely used in high-power applications due to its high power density, bidirectional power flow, and soft-switching capability. However, improving efficiency remains a major challenge. Existing strategies including mathematically-derived methods and artificial intelligence approaches still rely on complex and time-consuming analytical modeling or data-driven modeling, consequently increasing development complexity. To address these issues, this work proposes a model-free optimization framework based on deep reinforcement learning, enabling direct policy learning through system interaction without explicit modeling of electrical parameters, significantly reducing development time while ensuring optimization performance. To verify the generalization, a parameter sensitivity analysis is conducted, confirming strong generalization under different conditions of the converter. Furthermore, a degenerate variant of the deep deterministic policy gradient algorithm is developed for single-step decision optimization in 3p-DAB converters, along with an AI-driven duty-cycle control strategy for efficiency enhancement. Finally, comprehensive comparisons with state-of-the-art mathematical analytical method and data-driven approache validate the effectiveness of the proposed approach.