Dual-analysis framework for metro station classification: based on temporal flow pattern features

Abstract

Understanding passenger flow patterns is essential for flow prediction. This study proposes a dual classification framework for metro stations derived entirely from AFC data, without reliance on external sources. Manual classification defines three fundamental flow types: Bidirectional Tidal Flow (BTF), Sustained Outbound Flow (SOF) and High-Volume Flow (HVF), and quantifies their proportions within each station’s flow composition. Machine learning classification groups stations with similar overall flow profiles using multi-scale temporal features and attention mechanisms, identifying network-level similarities. Results show that: (1) BTF patterns dominate, with high weekday-weekend peak synchronization and stable proportional relationships; (2) the machine learning classifier achieves meaningful station groupings aligned with manual flow patterns; (3) classification-derived features from both components demonstrably improve downstream passenger flow prediction accuracy.