Deep learning models for urban rail transit traffic flow prediction
TL;DR
For the short-term passenger flow prediction task of urban rail transit, a deep learning model integrating temporal network, graph convolution, and attention mechanism is proposed that demonstrates higher accuracy and more stable generalization ability in the short-term passenger flow prediction task.
Abstract
For the short-term passenger flow prediction task of urban rail transit, a deep learning model integrating temporal network, graph convolution, and attention mechanism is proposed. The method is based on AFC passenger flow data and constructs supervision sample tensors at a 15-minute granularity. Time encoding, weather variables, and holiday markers are introduced to complete multi-source feature fusion. TCN/GRU is used to extract temporal dependencies, GCN is used to depict the topological propagation relationship of stations, and the attention mechanism is used to achieve weighted key feature selection. The model training uses sliding windows to generate samples, and combines the MSE loss function and Adam optimization to complete parameter updates. Experimental results show that the RMSE and MAPE of the proposed model reach 16.85 and 9.74% respectively, which are 12.51% and 18.43% lower than those of the GRU model, and 6.08% and 10.31% lower than those of the TCN-GCN model. It demonstrates higher accuracy and more stable generalization ability in the short-term passenger flow prediction task.