Abstract: Post-fault transient stability assessment and dynamic power angle prediction are of great significance for the stable operation and emergency control decision of power systems. Directly using all the initial transient response observations to predict power angle trajectories in the whole transient process will cause the input and output of the model to be too complex and difficult to mine the association rules, which restricts the utility in larger-scale power systems. Therefore, this paper proposes a dynamic power angle trajectory prediction and stability assessment method for power systems based on two-stage information compression. In stage one, the initial transient response is mapped into low-dimensional Euclidean space by uniform manifold approximation and projection（UMAP） algorithm to compute mutual information as correlation measurement. Then, feature selection is implemented by improved maximum relevance and minimum redundancy（im RMR） algorithm which synthesizes relevance and redundancy to eliminate redundant information from the response feature set. In stage two, a resolution adaptive largest triangle three buckets（RALTTB） algorithm is utilized to compress the power angle trajectories in the whole transient process to enhance the information density of the trajectory data. Finally, the dynamic compressed power angle trajectory prediction and stability assessment model is built based on gated recurrent unit（GRU） and the effectiveness is verified by CEPRI-TAS case system.