Graph transformer embedded deep learning for short-term passenger flow prediction in urban rail transit systems: A multi-gate mixture-of-experts model

Abstract

Urban rail transit (URT) plays a crucial role in mitigating urban traffic congestion by offering faster and higher-quality travel services. Short-term passenger flow predictions have practical significance for metro management and operation. However, the complex spatiotemporal characteristics and the relationship between entry and exit passenger flows make it challenging to detect the dynamic evolution patterns. This study proposes a Spatio-Temporal Graph Transformer (STGT) under the multi-task learning framework, utilizing Graph Transformer network and gated residual units to select and aggregate features. To account for the correlation between entry and exit passenger flow prediction tasks, the STGT model integrates a Multi-gate Mixture-of-Experts (MMoE) approach, which combines different expert networks for diverse input and explicitly learns to model passenger flow relationships in various scenarios. Metro-related characteristics such as weather conditions, train operation characteristics, and accessibility of nearby bus stops are incorporated to enhance prediction accuracy. Experimental evaluations are conducted using real-world historical passenger travel records from the Beijing subway. The results demonstrate the superior robustness and advantages of the STGT-MMoE model over basic and advanced benchmarks for passenger flow prediction tasks. These findings provide compelling evidence to address the challenges of short-term inflow and outflow prediction in urban rail transit systems.