IS15.7 - Lessons Learned Converting High-Fidelity Switching Models of Resonant Converters into AI-Based Reduced Order Models for System-Level Simulations

High-fidelity switching models of resonant power converters create computational bottlenecks for system-level analysis, with simulations requiring hours of computation time. AI-based reduced order models offer dramatic speedups, reducing computation from hours to seconds, but require systematic methodology for successful implementation.

This presentation shares practical lessons learned from developing AI ROMs for LLC resonant converters using Deep State Space Models. Through extensive investigation of architectures and training parameters, key principles have emerged: focusing on relevant dynamics, using effective excitation patterns, maintaining lean network structures, and optimizing dataset preprocessing. These guidelines address common pitfalls and enable power electronics engineers to achieve high-accuracy ROMs with orders-of-magnitude speed improvements for complex power system simulations.