Topology-Robust Power System Stability Prediction with a Supervised Contrastive Spatiotemporal Graph Convolutional Network

Modern power systems face growing challenges in stability assessment due to large-scale renewable energy integration and rapidly changing operating conditions. Data-driven approaches have emerged as promising solutions for real-time stability assessment, yet their performance often degrades under network topology reconfigurations. To address this limitation, the Spatiotemporal Contrastive Graph Convolutional Network (STCGCN) is proposed for the joint task prediction of voltage and transient stability across known and unknown topologies. The framework integrates a graph convolutional network (GCN) encoder to capture spatial dependencies and a temporal convolutional network to model electromechanical dynamics. It also employs supervised contrastive learning to extract discriminative features due to the grid topology variation, enhance stability class separability, and mitigate class imbalance under varying operating conditions, such as fluctuating loads and renewable integration. Case studies on the IEEE 39-bus system demonstrate that STCGCN achieves 89.66% accuracy on in-sample datasets from known topologies and 87.73% on out-of-sample datasets from unknown topologies, outperforming single-task learning approaches. These results highlight the method’s robustness to topology variations and its strong generalization across configurations, providing a topology-aware and resilient solution for real-time joint voltage and transient stability assessment in power systems.