Safety-Constrained Data-Driven Supervisory Scheduling of Virtual Inertia and Active-Power Release for a Three-Level NPC Grid-Forming Inverter

This paper develops a lightweight, safety-constrained supervisory scheduler for a switching three-level neutral-point-clamped (NPC) grid-forming (GFM) inverter. The scheduler maps grid strength, DC-source resistance, an AC active-power reference change, and a DC-load disturbance to two low-frequency commands: virtual inertia and an active-power release coefficient. Training labels are generated from closed-loop MATLAB/Simulink R2021b switching simulations and screened against six study-specific requirements covering frequency deviation, rate of change of frequency (RoCoF), root-mean-square (RMS) current, DC-link voltage, neutral-point voltage, and final active-power tracking. A standardized ridge-regression predictor is followed by an empirical feasible-set projection with an uncertainty-dependent inward margin. For a perturbed interpolation case, a fixed virtual synchronous generator (VSG), a tuned rule-adaptive scheduler, and unconstrained data-driven scheduling produce RoCoF values of 1.5156, 0.9975, and 0.6334 Hz/s, respectively, whereas the margin-aware scheduler achieves 0.4070 Hz/s and satisfies all six requirements. Under the strongest tested plant mismatch, increasing the virtual inertia to J = 8 and reducing the active-power release coefficient to r = 0.475 restores feasibility. Additional tests show that command feasibility depends on scheduler activation timing and command slew, while the projection maintains feasibility under deterministic feature-estimation bias. Under a severe power request outside the feasible operating envelope, conservative projection reduces the peak RMS current from 1.7248 p.u. for the fixed VSG to 1.0645 p.u., demonstrating preventive current-stress mitigation rather than hard current limiting. Overall, the results demonstrate preventive, constraint-aware supervisory scheduling across the tested switching-simulation envelope.