Grid Mirror: Harnessing Adversarial PINNs to Model Power Grid Dynamics

This paper presents a novel approach to estimate the dynamic behavior of power grids using Physics-Informed Neural Networks (PINNs) from an adversarial perspective. The methodology involves injecting a load perturbation into the grid to induce a response from the generators to estimate the grid’s dynamic response. The perturbations and generator frequencies are used to train a PINN, to estimate the grid’s state-space model by including it in the PINN’s loss function. By integrating physics constraints, the PINN ensures accurate and physically consistent estimations of grid dynamics. The PINN is initialized through educated guesses of the unknown matrices to ensure faster and more accurate conversion. The PINN demonstrates great potential with the estimated system states and generator frequencies closely following their actual counterparts with a correlation coefficient greater than 0.99. The achieved Relative Root Mean Square Error (RRMSE) is in the order of 1 x 10-4.