Abstract: For the data-driven power system transient stability assessment method, when the operation mode and topology of the power system is changing frequently, its application effect in the actual system is getting worse. To solve this problem, a selfadaptive transient stability assessment method based on the improved domain adversarial transfer learning is proposed. Based on the characteristics of the power system measurement data, a deep neural network is designed. After the operation scenario is changed, the gradient reversal layer is used to add the domain adversarial training mechanism to extract the common features between the source domain and the target domain, thus the distribution difference between the domains and the demand for training samples are reduced. At the same time, the knowledge of the source domain model is transferred, and the parameters of the feature extractor are updated synchronously to ensure the rapidity and accuracy of the model updating. The test results of the IEEE 39-bus system and the South Carolina 500-bus power grid in USA show that the proposed method can reduce the size of training samples in the target domain by reasonably transferring the original data and model, and has rapidity, versatility, and strong adaptability.