Abstract: At present, power system is moving towards a high proportion of renewable energy access, power electronics, and increasingly tighter interconnection, which puts forward higher requirements for the accuracy and real-time performance of transient stability assessment. Using a data-driven power system transient stability evaluation method can only utilize the dynamic response time series data after a fault to achieve real-time and accurate transient stability evaluation. This paper proposed a multi-stage power system transient stability evaluation method based on sample attention and hierarchical features to achieve real-time and accurate evaluation of transient stability. First, from the energy function point of view, time series data such as the original value of δ/V/θ/P/Q, integral quantity, and differential quantity were selected as the original input feature quantity. At the same time, to characterize the importance of the sample for stable rule learning, this paper defined sample attention index based on SVM pre-classification. Further, the use of reference load level information and stability labels to construct hierarchical feature learning supervision to enhanced the stability of feature extraction. Finally, based on the timing characteristics of LSTM's own output results, a multi-stage power system transient stability assessment scheme was proposed, which kept the error rate at a low level while ensuring a high classification accuracy rate. The test results of an IEEE 10-machine 39-bus system and a regional power grid verify the accuracy and necessity of the method in this paper.