With the rapid development of the power system, the large-scale integration of new energy into the grid and the coordinated optimization of source-grid-load-storage have increased the proportion of power electronic equipment, making the stability of the power grid, especially the assessment of transient stability, particularly important. Aiming at the problem of insufficient consideration of topological structure in traditional methods, a deep learning method based on Transformer-graph attention network(GAT) parallel feature fusion was proposed for the transient stability evaluation of power systems. The busbar voltage amplitude, phase angle and topological adjacentation matrix were taken as input features. Batch data were generated using the Siemens simulation software PSS/E, and features were extracted in parallel through Transformer and GAT. Weighted fusion was carried out using the attention mechanism. The comparison results with other methods show that this method simulates different load conditions and fault conditions in the IEEE 39-node system. The results indicate that the evaluation accuracy and robustness are relatively high, and it can effectively improve the safety and stability of the power system.