This paper proposes a novel transient stability assessment (TSA) method based on double-stage support vector machine (double-SVM). Firstly, an initial sample set composed of the system and single machine features is constructed, and the binary particle swarm optimization (BPSO) algorithm and the within-class and between-class distance criterion are used to search the optimal feature subset from the original sample set to reduce the feature dimension and eliminate redundant information. Then, based on the idea of double-SVM, the penalty loss caused by the label noise in the objective function of the second stage pin-SVM is reduced by the p-value generated by the first stage SVM, so that they will not play a leading role in the establishment of the hyperplane, and the effect of the label noise on classification results can be avoided while the model is also insensitive to the feature noise due to the characteristics of pin-SVM itself. Thus the model has better performance and robustness. Finally, the effectiveness and accuracy of the proposed method are verified by IEEE-39 system simulations.