Graph Attention Informer for Long-Term Traffic Flow Prediction under the Impact of Sports Events

Yaofeng Song, Ruikang Luo, Tianchen Zhou, Changgen Zhou, Rong Su2024 · DOI

TL;DR

This paper introduces the "Graph Attention Informer" (GAT-Informer), a novel deep learning model designed to improve long-term traffic flow prediction by explicitly incorporating the impact of large-scale sports events. By combining graph attention for spatial relationships and the Informer architecture for efficient long-sequence forecasting, GAT-Informer demonstrates superior accuracy, especially when accounting for event-driven traffic surges.

What Problem Does This Paper Address?

Accurate traffic flow prediction is crucial for intelligent transportation systems (ITS), aiding in traffic management, urban planning, and individual route optimization. While deep learning models have advanced short-term prediction, long-term forecasting remains challenging, particularly when external factors like major sports events cause significant, temporary disruptions. These events lead to unpredictable surges in traffic, complicating standard prediction models that often rely on historical patterns alone. Existing methods struggle to capture both the complex spatial dependencies of road networks and the dynamic temporal impact of events, leading to decreased accuracy for longer prediction horizons.

Methods

The proposed GAT-Informer model integrates two key components:

  • Graph Attention Layer (GAT): This layer leverages graph neural networks to model the road network as a graph. It uses attention mechanisms to dynamically learn the relationships between different road segments (nodes) and their connectivity (edges), capturing spatial dependencies more effectively than static graph structures. Crucially, it incorporates an "attention score" mechanism to weigh the importance of neighboring nodes, allowing it to adapt to changing traffic conditions.
  • Informer Layer: This component is an efficient variant of the Transformer architecture, specifically designed for long-sequence time-series forecasting. It addresses the computational complexity and memory demands of standard Transformers, making it suitable for predicting traffic flow over extended periods. The Informer layer uses a "ProbSparse" attention mechanism to efficiently capture long-range temporal dependencies.

A novel aspect is the introduction of a Sports Events Impact Factor. This factor is mathematically modeled using a Gaussian distribution to represent the temporal and spatial influence of sports events around an arena. This factor is then fused with the input data at both the GAT and Informer layers, allowing the model to explicitly learn and adapt to event-induced traffic patterns.

Datasets

The GAT-Informer model was evaluated on two datasets:

  • PeMS04 Dataset: A widely used benchmark dataset for traffic flow prediction, containing data from 307 detectors across 29 roads in the San Francisco Bay Area. For this dataset, the impact of sports events was disregarded to focus on general time-series prediction capabilities.
  • London Dataset: A newly collected real-world dataset featuring data from 96 detectors across three major motorways (M1, M4, M25) around London. This dataset specifically includes data from the period of the Laver Cup, a major sporting event, allowing for the evaluation of the model's ability to handle event impacts.

Key Insights and Ideas

  • Explicitly modeling event impact is crucial: The study demonstrates that simply relying on historical traffic data is insufficient for accurate long-term prediction when major events occur. The GAT-Informer's ability to incorporate a dedicated sports event impact factor significantly improves prediction accuracy.
  • Graph attention excels at spatial dynamics: The GAT layer effectively captures the complex spatial relationships within the road network, which is essential for understanding how traffic flows across different locations.
  • Informer's efficiency for long sequences: The Informer architecture proves vital for handling the long-term prediction horizon, maintaining performance where other models might falter due to computational constraints or decaying accuracy.
  • Event impact is both temporal and spatial: The model's formulation of the impact factor considers both the timing and geographical proximity to the event, reflecting real-world traffic behavior.

Key Results

  • Superior long-term prediction: On the London dataset, GAT-Informer consistently outperformed baseline models (Informer, ASTGCN) across various prediction horizons (up to 120 minutes). For instance, in the 120-minute horizon, GAT-Informer achieved lower RMSE, MAE, and MAPE compared to ASTGCN.
  • Impact of sports events: Removing the sports events impact factor from the GAT-Informer model led to a noticeable decrease in prediction accuracy across all horizons, validating the importance of this feature.
  • Robustness to prediction horizon: The GAT-Informer showed less performance degradation as the prediction horizon increased compared to other models, indicating its stability for long-term forecasting.
  • PeMS04 performance: Even without event data, GAT-Informer performed competitively on the PeMS04 dataset, often outperforming or matching advanced baseline models like ASTGNN, particularly for longer horizons.

Limitations and Future Work

  • Data Uncertainty: The authors acknowledge that real-world traffic data can have missing values due to detector failures or blackouts. Future work aims to address this by introducing fuzzy representation into the deep learning model to reduce the impact of data uncertainty and improve model adaptability to various datasets.
  • Redundancy parameter (Δt): While the study explored the impact of the Δt parameter (representing early arrival/delayed departure), the performance changes were not drastic. Further fine-tuning or adaptive methods for this parameter could be explored.
  • Generalizability: While tested on two distinct datasets, further validation on a wider range of urban environments and event types would strengthen the model's generalizability.

Summary updated

BibTeX
@misc{song2024graph,
  author = {Yaofeng Song and Ruikang Luo and Tianchen Zhou and Changgen Zhou and Rong Su},
  title = {{Graph Attention Informer for Long-Term Traffic Flow Prediction under the Impact of Sports Events}},
  year = {2024},
  doi = {10.3390/s24154796},
}
Figure