Abstract: An event-aligned multi-fidelity physics-informed neural operator is proposed to reduce the cost of multi-scenario dynamic assessment and improve predictions near disturbance transitions in grid-forming renewable power plants. A differentiable aggregate model with a single point of common coupling (PCC) represents the coupled dynamics of grid-forming and grid-following converters. An explicit solution of the PCC voltage links frequency, voltage, power and current responses through the network equations. Known disturbance information is transformed into local event coordinates, physical forcing and edge features for an initial-state-anchored residual decoder. Low-fidelity trajectories cover the parameter domain, while a small set of high-fidelity trajectories and nominal dynamic residuals constrain transient predictions. Equivalent inertia and damping are identified using a weighted-least-squares-assisted physics-informed neural network. With identical raw event information, the mean absolute errors of frequency nadir and voltage minimum on 32 in-domain test cases are 0.125±0.013 Hz and 0.0131±0.0014 p.u., respectively, approximately 64% and 38% lower than those of standard DeepONet. The frequency-nadir error on 42 weak-grid cases is 0.133±0.016 Hz. Capacity scans with fixed plant-level inertia and damping show that increasing the grid-forming share raises aggregate branch current but reduces peak current on the converters’ own capacity base. Current-limiting cases further establish a workflow that combines batch prediction with dynamic checks of near-limit conditions, supporting capacity allocation that accounts for both plant support performance and converter current headroom.