Abstract: The data-driven method has shown potential in the accuracy and timeliness of transient stability assessment (TSA). However, the limitation of its application lies in the high-dimensional characteristics of the power system, which leads to the long time-consuming training of the algorithm and the generalization performance of the single prediction model. It is difficult to cope with the complex and changeable power system operation scenarios and needs to be updated quickly. In order to solve this problem, this paper constructs a framework based on active transfer learning. Firstly, the basic model is built and trained based on the original scene data. The update mechanism is started when the performance of the model decreases due to the change of the running scene. A large number of samples without stable state are generated by short-term time-domain simulation, and a small batch of labeled samples are generated by complete simulation. The active learning method based on variational adversarial is used to learn the potential feature representation space of the data, and the unlabeled samples with the largest amount of information are selected and labeled according to the confidence. Finally, the basic model parameters are migrated and fine-tuned with labeled samples to save the update time while ensuring the migration accuracy. The IEEE 39 node verifies the effectiveness of the proposed method.