With the continuous expansion of power systems and the growing penetration of renewable energy, modern power grids have become increasingly complex. Local node failures can easily trigger cascading outages, posing serious threats to system security. Therefore, proactively identifying vulnerable nodes and implementing targeted protection measures is crucial for ensuring safe grid operation. To achieve efficient vulnerable node identification, an improved graph convolutional network-Transformer (GCN-Transformer) method is proposed. First, a set of node vulnerability evaluation indicators is constructed by integrating complex network theory with an improved information entropy–K-shell algorithm. Second, a Chebyshev polynomial-based Kolmogorov-Arnold Network (Cheb-KAN) is introduced as a front-end branch feature extraction module for the graph convolutional network (GCN), enhancing the propagation effectiveness of node feature extraction across different branches. The features extracted by the improved GCN are then fed into a Transformer integrated with a multimodal cross-attention (MCA) mechanism to capture global correlations among different modal features, thereby establishing a deep learning model for vulnerable node identification. Subsequently, multiple operating scenarios are constructed based on the IEEE 39-bus system to build the original dataset for model training. Finally, the proposed model is trained and evaluated on this dataset. Results demonstrate that the proposed method significantly outperforms traditional graph network models in terms of identification accuracy, showing strong feasibility and promising engineering application potential in practical power grid scenarios.