Structured neural ODE for nonlinear thermal reduced order models of power converters

Accurate thermal modeling is essential for the reliable design and operation of Power Electronic (PE) systems, particularly as nonlinear material properties significantly influence thermal dynamics. In this work, we propose a novel model reduction strategy based on structured Neural Ordinary Differential Equations (Neural ODEs) for the nonlinear thermal behavior of power inverters. Our method leverages a physicallyinformed architecture that approximates the nonlinear components of the system. The approach is validated on a nonlinear FEM model of a SiC-based inverter. Results demonstrate that the proposed Neural ODE-based Reduced Order Model (ROM) achieves higher accuracy than Proper Orthogonal Decomposition (POD)-Discrete Empirical Interpolation Method (DEIM), with significantly reduced computational cost, achieving up to 100× speed-up compared to implicit DEIM. Moreover, the resulting model can be seamlessly integrated into system-level simulation platforms, enabling efficient thermal management studies and control design. The proposed Neural ODE-based ROM can be efficiently implemented on modern microcontrollers, with measured inference times on the TC387 of 330µs per time step and minimal memory requirements, demonstrating its potential for real-time embedded thermal management applications.