Physics-aware dynamic graph embedding with contrastive feature alignment for transient stability prediction under grid topology variations

Accurate, low-latency online prediction of transient stability is essential for secure operation of modern power systems subject to disturbances. Although data-driven deep learning methods have shown strong predictive performance, their accuracy often deteriorates when system parameters or operating conditions change, particularly under grid topology variations. This lack of adaptability limits their effectiveness in real-world applications. By introducing a physics-informed inductive bias derived from multi-machine swing dynamics, this paper proposes a physics-aware Dynamic Graph Embedding (DGE) that encodes time-synchronized phasor measurement unit (PMU) signals together with network structural information into compact, node-wise representations, and a DGE-based Supervised Contrastive Learning (DGE-SCL) framework utilizing a lightweight Convolutional Neural Network (CNN) backbone. This framework combines topology-invariant data augmentation with supervised contrastive feature learning to obtain topology-robust, class-discriminative embeddings. These components are applied to real-time transient-stability classification, enabling efficient transfer of pretrained predictors across different network configurations. The method is evaluated on the Institute of Electrical and Electronics Engineers (IEEE) 39- and 145-bus test systems under multiple N-1 and N-m-1 topology scenarios; results show consistently improved generalization and robustness compared to baselines while maintaining low inference latency suitable for near-real-time deployment.