Abstract: Currently, transient voltage stability assessment models based on deep learning face the following three major challenges: firstly, a large amount of transient data is required to train the network model, and the process of collecting these data is extremely tedious and inefficient; secondly, the key information embedded in the time-series data of multiple features in the transient process is difficult to be captured efficiently; and lastly, the imbalance between the transient voltage stabilization and voltage destabilization samples leads to a significant bias in the training process of the transient voltage stability assessment model and may lead to serious misjudgment consequences. Voltage stability assessment model has obvious tendency in the training process and may trigger serious misjudgment consequences. To address the above problems, a data enhancement based on improved WGAN network and bidirectional gated cyclic unit assessment model based on the attention mechanism are proposed. In order to make deep learning applicable to small datasets, a generative adversarial network is introduced for data augmentation, which extends the original dataset by generating additional effective samples through the generative network, and a bidirectional gated cyclic unit evaluation model based on the attentional mechanism is developed in order to extract the time dependence from the dynamic trajectory after system perturbation, which learns significant time dependence by bi-directional learning and assigns the attentional weights automatically. Finally, the tendency and misclassification problems due to sample imbalance during the training process of the transient voltage stability assessment model are further mitigated by an improved focal loss function. The test results show that the proposed method can significantly improve the accuracy of the model when dealing with both the original small dataset as well as the sample imbalanced dataset, and has a certain degree of anti-noise ability.