Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Multi-Channel Spatial-Temporal Transformer Model for Traffic Flow Forecasting

Jianli Xiao, Baichao Long

TL;DR

The paper tackles robust traffic flow forecasting by addressing long-horizon degradation and complex spatio-temporal dependencies. It introduces MC-STTM, a multi-channel spatial-temporal Transformer that fuses short-term and daily patterns through a multi-channel mechanism, while a spatial block combines adaptive and fixed graph structures with a Transformer for temporal reasoning. Key innovations include learning a dynamic adjacency via $A_{adp}$ alongside distance-based fixed graphs, and a weekday-aware position encoding within the Temporal Transformer. Experiments on six real-world datasets show MC-STTM consistently outperforms state-of-the-art baselines, with strong long-term predictive performance and resilience to noise, suggesting practical value for urban traffic management and planning.

Abstract

Traffic flow forecasting is a crucial task in transportation management and planning. The main challenges for traffic flow forecasting are that (1) as the length of prediction time increases, the accuracy of prediction will decrease; (2) the predicted results greatly rely on the extraction of temporal and spatial dependencies from the road networks. To overcome the challenges mentioned above, we propose a multi-channel spatial-temporal transformer model for traffic flow forecasting, which improves the accuracy of the prediction by fusing results from different channels of traffic data. Our approach leverages graph convolutional network to extract spatial features from each channel while using a transformer-based architecture to capture temporal dependencies across channels. We introduce an adaptive adjacency matrix to overcome limitations in feature extraction from fixed topological structures. Experimental results on six real-world datasets demonstrate that introducing a multi-channel mechanism into the temporal model enhances performance and our proposed model outperforms state-of-the-art models in terms of accuracy.

A Multi-Channel Spatial-Temporal Transformer Model for Traffic Flow Forecasting

TL;DR

The paper tackles robust traffic flow forecasting by addressing long-horizon degradation and complex spatio-temporal dependencies. It introduces MC-STTM, a multi-channel spatial-temporal Transformer that fuses short-term and daily patterns through a multi-channel mechanism, while a spatial block combines adaptive and fixed graph structures with a Transformer for temporal reasoning. Key innovations include learning a dynamic adjacency via alongside distance-based fixed graphs, and a weekday-aware position encoding within the Temporal Transformer. Experiments on six real-world datasets show MC-STTM consistently outperforms state-of-the-art baselines, with strong long-term predictive performance and resilience to noise, suggesting practical value for urban traffic management and planning.

Abstract

Traffic flow forecasting is a crucial task in transportation management and planning. The main challenges for traffic flow forecasting are that (1) as the length of prediction time increases, the accuracy of prediction will decrease; (2) the predicted results greatly rely on the extraction of temporal and spatial dependencies from the road networks. To overcome the challenges mentioned above, we propose a multi-channel spatial-temporal transformer model for traffic flow forecasting, which improves the accuracy of the prediction by fusing results from different channels of traffic data. Our approach leverages graph convolutional network to extract spatial features from each channel while using a transformer-based architecture to capture temporal dependencies across channels. We introduce an adaptive adjacency matrix to overcome limitations in feature extraction from fixed topological structures. Experimental results on six real-world datasets demonstrate that introducing a multi-channel mechanism into the temporal model enhances performance and our proposed model outperforms state-of-the-art models in terms of accuracy.
Paper Structure (25 sections, 11 equations, 10 figures, 4 tables)

This paper contains 25 sections, 11 equations, 10 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Temporal Models
  4. Spatial Models
  5. Hybrid Models
  6. Proposed Model
  7. Problem Definition
  8. Model Framework
  9. Multi-Channel Mechanism
  10. Spatial-Temporal Block
  11. Feature Fusion
  12. Prediction Layer
  13. Multi-Channel Mechanism
  14. Spatial Block
  15. Temporal Transformer
  16. ...and 10 more sections

Figures (10)

  • Figure 1: Visualization of spatial-temporal correlation: (a) distribution of urban road and traffic sensors, where red, yellow, and blue dots represent traffic sensors at different locations with different traffic conditions; (b) time slices of road network from time $t_1$ to $t_N$, where the current traffic flow at each node is not only related to the traffic conditions in the previous slice of the day, but also related to the historical traffic flow at the same time on previous days.
  • Figure 2: The framework of MC-STTM.
  • Figure 3: The architrctaure of spatial-temporal block.
  • Figure 4: Visualization of sensor distribution in the METR-LA dataset.
  • Figure 5: MAE comparison of flow prediction before and after introducing multi-channel mechanism in temporal models: (a) results on PEMS03; (b) results on PEMS04; (c) results on PEMS07; (d) results on PEMS08.
  • ...and 5 more figures