NeurPecs: Physics-Informed AI-Based Adaptive Circuit Simulator for Power Converters

This is the authors' abstract. We don't add key points for this paper.

As renewable energy systems become more complex, power converters require rigorous verification through simulations before hardware prototyping. The prevailing simulators are typically categorized into offline and real-time types. While offline simulators are more cost-effective, they often suffer from long computational time, whereas real-time simulators are constrained by time-step and switching frequency limitations due to computational resources and communication delays. Moreover, both types of simulators rely on fully defined, fixed circuit parameters based on prior assumptions, making them prone to inaccuracies under fluctuating or uncertain conditions. To address these limitations, this article introduces NeurPecs, a neuralized, adaptive simulator for power converters. We propose a novel, physics-informed recurrent neural architecture that integrates two neural loops to model both circuit and control dynamics for multitimescale simulations, characterizing steady-state and transient behaviors. By leveraging neural representations and advanced runtime optimization, NeurPecs achieves high-speed simulation and adaptivity. The feasibility of NeurPecs is demonstrated through three case studies: a dual-active-bridge, a buck and a flyback converter. The simulator delivers a mean average percentage error of 1.4% and operates 9.2x times faster than the PLECS simulator when deployed on resource-constrained edge platforms. This work aims to pioneer the next generation of physics-informed AI-based circuit simulators for power converters, providing high-speed, more adaptive, and cost-effective solutions.