Neural CDEs Model Grid-Forming Inverters for Power System Simulation

The application of artificial intelligence in modeling power electronic converters is expanding, yet it faces hurdles such as difficulties in analyzing across multiple time scales and a lack of physics-aware evaluation criteria, which can lead to suboptimal performance. To address these issues, this research introduces a Neural Controlled Differential Equation (Neural CDE) framework. This framework is designed to learn continuous-time surrogate models specifically for grid-forming inverters, facilitating electromagnetic transient (EMT) simulations. A key advantage of this approach is its ability to relax the constraint of fixed sampling rates, thereby enabling multi-time-scale control analysis. The framework further proposes an affine-control formulation that incorporates dual slow/fast pathways, allowing it to capture the hierarchical and multi-scale dynamics inherent in converter behavior. To enhance stability and coherence, a physics-inspired regularization method is also employed. When evaluated against EMT-generated trajectories, the Neural CDE model accurately reproduced transient responses, maintained effective damping and dominant oscillatory characteristics, and demonstrated stable long-horizon rollouts. These results indicate that Neural CDE-based component modeling offers a physically consistent and robust surrogate modeling approach for detailed EMT-level simulation studies in power systems.