Abstract: The artificial intelligence method has made considerable achievements in power system transient stability assessment (TSA). A conventional deep network is generally regarded as a "black box" model, which limits the dependability of intelligent algorithms for practical engineering applications. Furthermore, conventional methods have limited ability to capture the dynamic evolution process of the power system. To solve the above problems, this paper introduces a multi-stage transient stability assessment method based on Transformer encoder, and missing alarms can be effectively reduced with multi-stage predictions. Compared with conventional methods, the Transformer model presents good interpretability, whose attention structure visualizes the internal work of the deep neural networks, so the model can adaptively identify and focus on the important features. Meanwhile, this paper utilizes multi-moment information to construct feature spaces. Compared with other networks, the Transformer gains a global receptive field via attention mechanism, and thus the state change of power system can be characterized quickly and accurately. Simulation on IEEE-39 bus system shows that the proposed method presents good interpretability compared to common data-driven models, showing higher accuracy of transient stability assessment and stronger robustness under data pollution.