Currently, in the field of transient stability assessment and control represented by Transformer, which is based on deep learning, there is an insufficiency in the ability to extract temporal information from power system transient data, and it is challenging to balance the speed and accuracy of the evaluation. This article proposes a two-stage transient stability assessment method based on Swin Transformer. This method replaces the self-attention module in standard Transformers by using non-overlapping local windows and shifted windows to calculate self-attention, which can more effectively extract temporal information from transient data and improve model evaluation accuracy. In addition, by analyzing the correspondence between attention weights and system instability patterns, the model decisions were explained, which improved the interpretability of the model. The simulation example of the IEEE 10-machine 39-node system shows that the proposed method has better performance evaluation compared to traditional deep learning and machine learning methods, and has lower computational cost compared to the standard Transformer.