Critical sample augmentation and adaptive spatio-temporal fusion network for transient voltage stability assessment

Accurate transient voltage stability assessment (TVSA) is hindered by the scarcity of unstable samples, the underutilization of critical boundary information, and limited spatio-temporal feature extraction capability of existing models. This paper proposes a TVSA framework that applies a least-squares GAN (LSGAN) network to augment critical samples, and an Adaptive Hybrid Spatio-Temporal Fusion Network (AHSTF-Net) for TVSA. First, a golden-section search method is used to determine the critical clearing time (CCT), thereby constructing voltage stability boundaries of the system. To enrich training data in the vicinity of stability boundary, a structure-preserving critical sample generation scheme (SP-CSG) is developed, where a least-squares GAN (LSGAN) is developed to generate a large number of critical samples, and a Structural Similarity Index Measure (SSIM)-based technique is used to select structurally and physically plausible samples among them, enhancing both the quality and diversity of the augmented data. Second, the AHSTF-Net, which integrates Transformer modules, deformable convolutions and residual connections, is developed to capture multi-scale spatio-temporal dependencies while improving training stability. Case studies on the IEEE 9-bus and 39-bus systems show that the proposed method substantially outperforms benchmark models in terms of data quality, diversity and classification accuracy. Even under strong noise conditions (20 dB), the method maintains an accuracy of TVSA above 99%, demonstrating its robustness and potential for practical engineering applications.