<p>Nodal displacement sequences of transmission towers display high spatial coherence with changes in node positions. However, existing prediction models frequently neglect the relationship between temporal and spatial, resulting in suboptimal prediction performance. Therefore, we proposed TS-Net, a novel short-term state prediction model for transmission towers designed to integrate the spatial–temporal features of the tower nodes. Temporally, the model incorporates Time2Vec to automatically capture essential periodic information and employs multi-head self-attention (MSA) to ascertain temporal associations. Spatially, the adjacency matrix is constructed based on the physical locations and dynamic time warping (DTW) analyses of key nodes, with the graph attention network (GAT) extracting spatial features from this matrix. For effective spatial–temporal connection establishment, we design an adaptive feature fusion module, cleverly introducing a "guide vector" to allocate spatial information to each time step, thereby achieving deep integration of spatial–temporal features. To verify the superiority of the proposed model, comparison and ablation experiments are carried out. The results indicate that incorporating spatial features significantly enhances prediction ability, exhibiting a mean square error of 0.0218%, a mean absolute error of 0.0864%, and a coefficient of determination of 96.96%. This paper provides a practical and accurate approach to predicting the state of transmission towers.</p>