Physics-Informed Neuro-Symbolic Control for 100% Inverter-Based Hybrid Microgrids

Abstract

The displacement of synchronous machines by inverter-based resources (IBRs) eliminates physical rotational inertia, exposing modern power grids to severe frequency and voltage instability during electromagnetic transients. Existing model-free reinforcement learning controllers fail to guarantee asymptotic stability or Lyapunov compliance under unlearned topological perturbations. This paper presents a Physics-Informed Neuro-Symbolic Reinforcement Learning (PINSRL) architecture for sub-cycle control of 100% IBR microgrids. By embedding differential-algebraic power balance equations and swing-equation dynamics directly into the temporal difference loss functional, the controller guarantees bounded state trajectories. Numerical simulations on a 33-bus hybrid feeder demonstrate 14 ms fault recovery and a 74.2% reduction in Rate of Change of Frequency (RoCoF) relative to the virtual synchronous machine baseline.