Temporal Contrastive Learning for Transient Voltage Stability Assessment

This paper proposes a novel temporal contrastive learning framework for transient voltage stability assessment in modern power systems characterized by high renewable integration and power electronics. Addressing the limitations of conventional methods—such as computational inefficiency in time-domain simulations, reliance on empirical thresholds in engineering criteria, and dependency on labeled data in supervised learning—our framework leverages unsupervised contrastive pretraining to extract discriminative temporal-contextual features from voltage trajectories. By integrating a dual-level augmentation strategy, cross-view temporal prediction, and contextual contrasting, the method disentangles noise-invariant stability patterns from transient artifacts. A Maximum Lyapunov Exponent (MLE)-based sample calibration strategy further refines model training by excluding ambiguous cases near stability boundaries. Evaluated on an IEEE 9-bus system, the proposed framework outperforms supervised learning and physics-based metrics. The strong correlation between SC and accuracy validates its ability to generalize in label-scarce scenarios, offering a scalable solution for real-time grid resilience monitoring.