Abstract: When the deep learning network is used to realize the transient stability assessment of the power system, the classification performance of the assessment algorithm is greatly affected due to the insufficient sample diversity and the poor anti-interference. In view of the above problems, this paper proposes a transient stability assessment method based on the double generator LSTM-generative adversarial network (DGL-GAN). In this method, on the one hand, the batch sample generator and the discriminator form an adversarial network which generates new samples in batches that match the true distribution through alternate training to learn the distribution characteristics of transient data, thus solving the problem of insufficient sample diversity in power system transient assessment. On the other hand, the repair generator has an LSTM autoencoder which can not only remove the noise in the transient data of the power system but also compensate for the missing fragments in simulation or measurement, solving the problem of poor anti-interference ability of the evaluation algorithm. In addition, the network structure design proposed in this paper is based on multi-layer LSTM which can further improve the model's feature extraction ability of the transient time series data. The simulation results in the New England IEEE 39-bus system show that the transient stability assessment model proposed in this paper can effectively enhance the sample diversity, significantly improve the transient stability assessment performance, and also have good anti-interference ability.