Forecasting Short-Term Passenger Flow via CBGC-SCI: An In-Depth Comparative Study on Shenzhen Metro

Abstract

Accurate short-term forecasts of passenger flow are crucial for the organization and management of metro systems. However, passenger flow in metro networks exhibits unique and complex patterns due to temporal dependencies associated with operating periods, graph-based spatial dependencies among distant stations, and correlation between inflow and outflow. To address these challenges, this paper proposes a novel deep learning approach, Chebyshev Graph Convolutional-Sample Convolution and Interaction Network (CBGC-SCI), for forecasting metro passenger flow. CBGC-SCI comprises two components: a spatial module based on Chebyshev polynomials graph convolution and a temporal module based on a downsample-convolve-interact architecture. The spatial module effectively captures graph-structured adjacency relations among metro stations, particularly those at a distance, by aggregating high-order node information. The temporal module extracts nonlinear passenger flow dynamics and long-term temporal dependencies, offering multi-step forecasting results with high predictive accuracy. Moreover, using smart card data of Shenzhen Metro collected via the Automatic Fare Collection (AFC) systems, we conduct extensive comparative experiments for methods validation, involving different input time steps, the correlation between passenger inflow and outflow, various operating periods, and diverse Graph Neural Network (GNN) modules. The results show that CBGC-SCI can well capture distant graph-based spatial dependencies and temporal dependencies, and outperform state-of-the-art baselines in terms of forecasting accuracy. Besides, considering the correlation between passenger inflow and outflow enhances the model's forecasting accuracy. Considering various metro passenger flow patterns during different operating periods, our comparative analysis of forecasting results highlights the superior performance of our proposed model across all operational conditions.