Dual-Orthogonal Defense: Integrating Physics Verification and Manifold Analysis for Robust Transient Stability Assessment

The transition of contemporary power systems toward low-inertia and highly nonlinear operational modes introduces stringent real-time requirements for transient stability assessment (TSA). Although deep learning (DL)-based TSA approaches have alleviated certain computational limitations, their intrinsic “black-box” characteristics and susceptibility to adversarial perturbations in high-dimensional feature spaces expose them to potential attacks. The paper presents a dynamic defense framework that integrates physical mechanism verification with manifold feature analysis. A dual-orthogonal architecture is established: the physical mechanism verifier (PMV), employing discretized swing equations (DSE), functions as a stringent constraint to preclude explicit breaches of system evolution principles; meanwhile, the data feature analyzer (DFA), built upon spatio-temporal graph attention networks (ST-GAT), captures non-Euclidean topological features and utilizes a physical residual embedding scheme to detect latent statistical irregularities. Furthermore, a physics-informed neural network (PINN) is deployed for online optimization during test time. Through the joint minimization of physical residuals and data fidelity loss, compromised measurement trajectories are projected back onto the physically admissible manifold. Simulation studies on the IEEE 39-bus system indicate that the proposed methodology enhances the detection rate for physics-aware attacks while sustaining millisecond-level online responsiveness.