Abstract: Overdependence on manual experience frequently leads to the irrational selection of orepass. Therefore, a scheduling model needs to be established to make sound decisions on orepass selection, increase the efficiency of underground rail transport, and improve production efficiency in metal mines. In this study, the −495 level of the Baixiangshan iron mine in Anhui Province is used as a research object. Orepass information, tunnel information, and historical mining, loading, and transporting data are collected. The data are then preprocessed to obtain a 130-order matrix that can describe the rail transit topology. Several vectors containing road/orepass basic information, road/orepass trajectory information, and orepass chronological material-level information are used for model training and validation. Time-series transformer graph convolutional network, which is denoted as T-TransGCN, is a temporal graph neural network that integrates orepass features, road features, and rail topology information. T-TransGCN is proposed to determine the optimal orepass selection. It enhances performance through splitting temporal features, fine-tuning the pooling layer architecture, and embedding edge features. Validated results show that (1) the T-TransGCN model is better than the Time-series multi-layer perceptron (T-MLP) and the Time-series graph convolutional network (T-GCN). The label accuracy, F1 score, and Top-3 accuracy of T-TransGCN improve by 7.33%, 17.00%, and 14.26% compared with those of T-MLP, which indicates that T-TransGCN can effectively integrate node attributes and topology information. Moreover, T-TransGCN has a relatively higher number of model parameters, more complex model structure, greater stability, and stronger fitting capability than T-GCN. (2) The addition of chronological material-level features to T-TransGCN increases its F1 score and Top-3 accuracy by 11.75% and 17.02%, while the addition of trajectory features improves them by 11.83% and 10.01%. Both new data preprocessing methods are effective in enhancing the generalization ability of T-TransGCN. The chronological material-level features help T-TransGCN understand the recent state of orepass, while the trajectory features reflect the importance of different orepasses dynamically. The trajectory features help the model understand structural information, such as the similarity of adjacent nodes or the similarity of forked nodes. (3) The addition of edge features further distinguishes orepass nodes from road nodes. The optimization of the outputs of the pooling layer helps avoid the distraction of unimportant information. When chronological features are split, the F1 score and Top-3 accuracy of T-TransGCN improve by 15.94% and 12.34%. This increment enhances the focus of the model on the chronological material-level information. The integration of the abovementioned model improvements further increases the fitting capability, generalization ability, and stability of T-TransGCN.