PE-GPT: A New Paradigm for Power Electronics Design
Fanfan Lin, Xinze Li, Weihao Lei, Juan J. Rodríguez-Andina, Josep M. Guerrero, Changyun Wen, Xin Zhang, Hao Ma
Abstract
Large language models (LLMs) have shown exciting potential in powering the growth of many industries, yet their adoption in the power electronics (PE) sector is hindered by a lack of specialized PE technical expertise and challenges in processing PE-specific data. This study presents a pioneering approach to establish a multimodal LLM tailored for PE design applications, named PE-GPT. The methodology involves enhancing PE-GPT with retrieval augmented generation from a PE knowledge base, and proposes a hybrid framework that integrates an LLM agent with metaheuristic algorithms, Model Zoo, and Simulation Repository. This enhances its multimodal processing capabilities and enables integration into the existing design workflow. The PE-GPT methodology is demonstrated with two case studies: modulation design of the dual-active bridge (DAB) converter and circuit parameter design of the buck converter. PE-GPT demonstrates a 22.2% increase in correctness compared to human experts. Against other leading LLMs, PE-GPT shows a 35.6% improvement in correctness and a 15.4% enhancement in consistency, reducing hallucination. Hardware experiments validate PE-GPT’s multimodal capabilities in optimizing a five-degree-of-freedom modulation strategy for the DAB converter. The generalization of PE-GPT to other PE design applications and associated AI ethical considerations are also discussed. This research concludes by outlining inspiring future research directions, encouraging researchers to expand the boundaries of the PE industry and advance toward a more intelligent era.