The actual large-scale power system at the provincial level or above has numerous nodes, and the input feature space of deep learning models used for transient stability assessment of power grids faces the curse of dimensionality. The high training cost and difficulty in ensuring generalization ability have become a bottleneck for the application of such methods in practical large-scale systems. A graph attention deep learning transient power angle stability evaluation model is proposed to address this issue, which is oriented towards stability control verification and utilizes network equivalence for graph dimensionality reduction. Firstly, a whole grid generators diagram is established, and node-similarity is constructed based on key parameters such as rotor inertias that determine transient power angle stability, and node similarity is used to improve the edge weights of the generator diagram. As a domain knowledge embedding method, the dynamic equivalence approach of the power grid is used for reference, the network parts outside the area involved in the study of stability control systems are partitioned based on generators diagrams and hierarchical clustering algorithms. The formed partitions are corresponding to equivalent nodes to form a reduced dimensional generators diagram, and a mapping of the original graph to the reduced dimensional graph nodes and edge weight parameters is established to achieve dimensionality reduction of the original input space. Finally, a graph attention deep learning model is established using the dimensionality reduction graph as input to achieve transient power angle stability evaluation of complex networks. The effectiveness and accuracy of the model are verified through comparative analysis on a practical stability control system example in China Southern Power Grid.