Introduction
Accurate traffic forecasting is essential for intelligent transportation systems, yet it remains challenging due to complex, non-linear spatio-temporal dependencies. While modern Graph Convolutional Networks (GCNs) effectively capture spatial interactions, they often fail to account for exogenous shocks such as road maintenance and construction. These "maintenance downtimes" create significant disruptions that ripple through the network. Most existing models treat traffic as a closed system of speeds and volumes, overlooking critical external modalities like lane closures. This paper introduces GCN-RWZ, a novel architecture designed to bridge this gap by integrating construction data as a primary input channel.
Methodology
The proposed Graph Convolutional Network for Roadway Work Zones (GCN-RWZ) leverages several innovations to handle multi-modal traffic data:
- Multi-Channel Feature Maps: The model processes traffic speeds ($X^S$) and construction data ($X^C$) through separate channels.
- Influence Zone Function: To mitigate the numerical sparsity of binary construction indicators, the authors employ a continuous function that calculates the impact of a work zone on neighboring nodes based on geometric distance, controlled by a decay hyperparameter $\lambda$.
- Learnable Fusion: Speed and construction channels are fused into a single "speed wave" using a learnable weight matrix and the Hadamard product.
- Spatio-Temporal Attention: Multi-head attention mechanisms allow the model to dynamically weight the relevance of spatial neighbors and historical time steps.
- Hybrid Architecture: The attention-based GCN outputs are processed by a Bi-directional RNN to capture complex temporal dynamics across the graph.
Datasets and Experimental Setup
The model was evaluated on three primary datasets:
- Tyson’s Corner: A new benchmark covering 131 mile-segments in Northern Virginia (2019), featuring synchronized speed and construction data.
- Los-loop & PEMS-BAY: Standard benchmarks used to test the model's underlying architecture (GCN-RWZ⁻) in the absence of construction data.
Key Results
GCN-RWZ consistently outperformed state-of-the-art (SOTA) baselines, including T-GCN, STGCN, and GraphWaveNet. On the Tyson’s Corner dataset for a 60-minute forecast, GCN-RWZ achieved an MAE of 3.49 and an RMSE of 5.86, notably better than ASTGCN (MAE 3.73, RMSE 6.00).
Even on datasets without construction information, the GCN-RWZ⁻ variant remained highly competitive, achieving an RMSE of 5.92 on the Los-loop dataset (30-min forecast) compared to 6.15 for ASTGCN. Ablation studies further confirmed that the learnable fusion method is superior to simple additive or multiplicative approaches.
Discussion and Insights
A critical finding is that treating construction as a binary variable is mathematically inefficient due to sparsity. The use of a decaying "influence map" allows the network to learn the spatial ripple effects of roadwork more effectively. Furthermore, the multi-head attention mechanism serves as a robust filter for long-horizon predictions, preventing the performance degradation typically seen as the forecast window expands.
While GCN-RWZ represents a significant step forward, future work will focus on integrating accident data—which was excluded in this study to isolate maintenance impacts—and addressing the computational scalability of multi-head attention for massive city-wide networks.
Summary updated