A supervised contrastive learning method with novel data augmentation for transient stability assessment considering sample imbalance

Fast and accurate transient stability assessment (TSA) is crucial for power system safety, as it provides the basis for operators to decide on emergency control actions. However, data-driven TSA methods are restricted by sample imbalance in real-world systems, characterized by a predominance of stable over unstable samples. Consequently, data-driven TSA methods may bias assessment rules towards classifying unstable as stable, potentially causing significant economic losses. To address this issue, we propose balanced supervised contrastive learning (BSCL) with a novel data augmentation method. First, the postfault transient response observations are utilized as augmented samples to enable the model to learn postfault features. Subsequently, a specially designed balanced supervised contrastive loss is proposed. In the contrastive learning process, the augmented samples belonging to the same class are pulled together in the embedding space, while simultaneously pushing apart samples from different classes. By doing so, the proposed BSCL method can learn future information and separate different classes. After contrastive learning, a classifier is connected to the trained model and undergoes further training for TSA, enabling the model to predict transient stability. Case studies conducted on an IEEE 39-bus system and an actual power system demonstrate that the proposed BSCL method achieves high identification accuracy in imbalanced datasets and possesses strong generalization capabilities and robustness. Furthermore, the results indicate that the proposed data augmentation method outperforms other data augmentation methods utilized in BSCL.