Physics-Embedded Neural ODEs for Sim-to-Real Edge Digital Twins of Hybrid Power Electronics Systems

Edge digital twins (EDTs) deployed around power electronics systems can significantly enhance the monitoring and control of the systems. However, they are usually resource-constrained but require high fidelity between digital models and real systems. With limited computing resources, existing event-or-data-driven modeling methods implemented on EDTs fail to comprehensively capture hybrid dynamics (i.e., continuous modes and mode transitions), leading to safety risks. To address these challenges, this article proposes a physics-embedded neural ordinary differential equation (PENODE) framework that includes a multimode event automata to model continuous operation for each mode and neural networks with prior physics knowledge to model discrete mode switching with great accuracy. A deployment workflow based on the framework is designed to provide an effective guideline to transfer digital twins (DTs) from cloud to edge. The proposed PENODE framework has been implemented on the EDT of a high-frequency dc–dc converter. Experimental results verify that PENODE achieves significantly higher accuracy and largely reduced resources across white-box, gray-box, and black-box benchmarks, validating its physical interpretability, efficient edge deployment, and real-time control enhancement.