STAR: One-Stop Optimization for Dual-Active-Bridge Converter With Robustness to Operational Diversity

In the fast-evolving landscape of artificial intelligence (AI) technology, significant strides have been taken to streamline and automate the design of power converters. However, prevalent limitations in existing AI-aided design approaches underscore the need for innovative solutions. Firstly, a lack of robustness to operational diversity hampers adaptability, requiring a complete AI model building from scratch when conditions, modulation strategies, or performance metrics for evaluation change. Secondly, the challenging trade-off between dataset size and design accuracy prolongs data acquisition through hardware experiments. Thirdly, the industry faces reluctance to adopt pure data-driven approaches due to their inherent lack of fundamental and explainable circuit insights. Addressing these challenges, this paper introduces a one-stop optimization methodology, the STAR methodology, tailored for designing modulation parameters for a three-level neutral-point-clamped converter. Leveraging a physics-in-architecture recurrent neural network (PA-RNN), STAR emerges as a robust solution, demonstrating resilience across diverse operating scenarios, accommodating multiple modulation strategies, and evaluating various performance metrics, including current stress, zero voltage switching, zero current switching, and more. The paper provides diverse design cases as user-friendly guidance and validates the methodology’s effectiveness through rigorous hardware experiments. This pioneering approach not only overcomes existing limitations but also paves the way for a more streamlined and adaptable future in power converter design.