Power systems are moving from big spinning generators to many inverter‑based resources (IBRs). That change makes it harder to predict whether the grid will be stable after small disturbances. This paper builds a physics‑informed neural network (PINN) that learns the small‑signal stability of high‑dimensional multi‑inverter systems across different operating conditions.

Instead of measuring impedance at many frequencies (frequency scanning), the authors train their PINN on short time‑domain step‑response data (SRD) produced by electromagnetic transient (EMT) simulations. A single 1‑second window per operating point was used in training. The network is modular: one part predicts the poles (the system’s eigenvalues, i.e., oscillatory modes) and another predicts the residues (which show how strongly each mode appears in each bus). They also use a Sobol‑sequence sampling method to pick operating points that cover the high‑dimensional space efficiently. The paper reports training on 1,000 operating points using SRD.

Because the PINN predicts poles and residues of the whole‑system admittance (the matrix that links small voltage perturbations to currents), its outputs have direct physical meaning. The model can map how these modes move when power flows change. That gives operators a way to see where oscillations might appear, to identify likely root causes, and to choose generation distributions that avoid risky modes. The authors say the trained PINN can generate these stability views in seconds, rather than running many costly EMT runs or frequency scans.

The method is demonstrated on small test systems. The paper reports full validation on a two‑IBR system and a four‑IBR system. In those cases the PINN reproduced the eigenvalue loci (the paths of modes as operating conditions change) and estimated the possible range of oscillatory modes under given power flows. These are results that are hard to obtain with conventional analytical tools or brute‑force EMT scans.