A short-term metro passenger-flow forecasting method based on data preprocessing and multi-factor spatial correlation

Accurate modeling of spatial dependencies is essential for short-term metro passenger-flow forecasting. However, relationships other than physical adjacency are difficult to represent in a graph and to incorporate into graph convolution. This paper therefore proposes BC-MGCN-DEGRU, a neural forecasting model that integrates data preprocessing, multi-factor spatial learning, and dynamic temporal feature extraction. The spatial module constructs three subgraphs representing physical adjacency, station-flow correlation, and station-location correlation. DeepWalk extracts high-order node embeddings from the station-flow and station-location subgraphs. These graph-specific representations are combined at the node level and reweighted by an attention mechanism before graph convolution. A BiLSTM-CBLOF preprocessing module detects and removes anomalous records by exploiting temporal correlations in passenger flow. To improve real-time performance and responsiveness to abrupt changes, recent passenger-flow observations are dynamically introduced into a GRU module enhanced by an extreme learning machine. The experimental results show that the proposed preprocessing, spatial, and temporal modules improve forecasting performance. BC-MGCN-DEGRU also achieves lower MAE, RMSE, and MRE values than the six baseline models considered in the experiments.