Abstract: At present, data-driven deep learning algorithm is widely used in the field of power system transient analysis, but in practice, there are few transient instability in the measured data samples of power system, and insufficient attention is paid to the instability samples, resulting in the weak anti-interference and generalization ability of the classification model, which affects the evaluation performance of the whole model. To solve this problem, a power system transient stability assessment method based on multi-granularity neighborhood rough set and improved bi-directional long-short-term memory network(IM-Bi-LSTM) is proposed. Firstly, the neighborhood rough set is used to find the optimal reduction subset at different granularity levels, and then the BiLSTM neural network is used to extract the timing information of the feature subset, and the attention model is introduced into the model to add more weights to the features related to the unstable samples. Through the focus loss function, the weight coefficient is introduced to adjust the tendency of model training, solve the imbalance between unstable samples and transient stability samples, and improve the evaluation performance of the model. The experimental results on the IEEE 10 machine 39 bus system show that compared with other algorithms, the proposed method has better classification accuracy and more stable results.