.png){kind=icon}
.png){kind=icon}
Return
Short-term metro flow forecasting via spatiotemporal joint self-attention and multi-graph weighted fusion
.png){kind=icon}
OA
AI
Q
J
Z
X
F
DOI：10.1016/j.ijtst.2025.11.001.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Short-term passenger flow forecasting is crucial for intelligent scheduling in urban rail transit, yet remains challenging due to complex spatiotemporal dynamics and heterogeneous station relations. While prior methods leverage graph neural networks (GNNs) and attention mechanisms, they often rely on decoupled spatial-temporal modeling and static graph fusion, limiting their capacity to capture real-world mobility dynamics. To address these issues, this paper proposes a unified and parameter-efficient deep learning framework, spatio-temporal self-attention with graph fusion (STSA-GF). STSA-GF integrates a spatio-temporal joint attention mechanism that concurrently models temporal dependencies and spatial interactions within a unified attention pathway. In addition, it introduces a gated fusion strategy that dynamically integrates three heterogeneous spatial graphs: metro topology, flow similarity, and origin-destination (OD) correlations. A long–short term graph co-modeling strategy is further incorporated to enhance the representation of periodic and disturbance-prone patterns. Extensive experiments on two real-world metro datasets from Shanghai and Hangzhou, China, demonstrate that STSA-GF consistently outperforms state-of-the-art baselines across multiple forecasting horizons. These findings affirm the theoretical value of STSA-GF in unifying spatiotemporal attention and multi-graph fusion within a cohesive architecture, offering a principled approach to modeling dynamic mobility patterns and heterogeneous spatial structures in urban rail systems.
Keywords:
Short-term passenger flow forecasting
Spatio-temporal attention
Multi-graph fusion
Long–short term dependency
Intelligent transportation systems
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
4.8
Papers:
1.5K
Citations:
1.5K
Organization
Cited Papers