Aiming at the phenomenon that the instability mode of modern interconnected power grid is no longer single and multiple swing instability occurs frequently after disturbance, a transient stability assessment method based on eXtreme Gradient Boosting-deep forest is proposed in this paper. An artificial feature set is constructed by using the bus voltage track cluster, and supervised feature coding was performed on the feature set by XGboost. The sparse matrix after supervised coding was tri-classified by deep forest, and the mapping relationship between large-scale data set and unstable mode was established. The simulation analysis is carried out on IEEE 39 and IEEE 140 nodes. The proposed method has high accuracy and anti-noise performance, and can effectively reduce the misjudgment rate of multiple swing instability. Moreover, it still has strong robustness when the synchronous phasor measurement unit is missing.