Abstract
Accurate traffic flow forecasting is essential for urban management, yet traditional deep learning models often fail to account for non-recurring congestion caused by work zones. This paper presents GCN-RWZ, a novel Graph Convolutional Network architecture that treats construction data as a primary input modality. By integrating hypergraph representations with a "speed wave" fusion mechanism, GCN-RWZ captures the complex spat-temporal disruptions caused by roadwork. Evaluations on real-world datasets demonstrate that GCN-RWZ significantly reduces prediction error during work zone events compared to state-of-the-art baselines.
Introduction
Traffic congestion represents a significant economic burden. While Graph Convolutional Networks (GCNs) have advanced the state-of-the-art in predicting "normal" traffic patterns, they frequently struggle with anomalies. Work zones, though infrequent compared to daily commutes, are critical events for city engineers and commuters alike. Standard models often dismiss the resulting slowdowns as statistical noise because they rely primarily on historical speed data. This research bridges the gap by making work zone information a first-class citizen within the forecasting architecture.
Methodology: The GCN-RWZ Architecture
The proposed GCN-RWZ (Graph Convolutional Network for Roadway Work Zones) model builds upon the Attention-based Spatio-Temporal Graph Convolutional Network (ASTGCN) framework. It introduces several key innovations to handle the unique characteristics of roadwork disruptions.
Hypergraph Representation
Unlike traditional graphs where roads are represented as simple connected points, GCN-RWZ employs a hypergraph. This structure allows the model to capture higher-order relationships, modeling how a disruption on a single segment impacts a broader "neighborhood" of related road segments more effectively than a standard adjacency matrix.
"Speed Wave" Fusion
The core novelty of the model is the "speed wave" feature map. Rather than using a binary flag for construction, the model fuses real-time speed data with construction data using a learnable weight matrix and Hadamard products. This approach treats construction as a modulating force on traffic flow, allowing the model to learn the specific degree of influence a work zone exerts on traffic states.
Multi-Feature Mapping and Attention
The model ingests five distinct feature maps: real-time speed, historical average speed, a "diff map" (delta between current and historical speeds), a time matrix for periodicity, and a construction map based on radial basis functions. A multi-head spatio-temporal attention mechanism then identifies the most relevant spatial and temporal inputs for final prediction.
Experimental Setup
The model was evaluated using two extensive real-world datasets from Virginia, covering the entirety of 2019. These datasets provide a robust testing ground for forecasting under disruptive conditions.
As shown in the table above, the Richmond dataset includes 95 road segments sampled every 15 minutes, while the Tyson’s Corner dataset offers higher temporal resolution with a 5-minute sampling rate across 131 segments.
Results and Discussion
GCN-RWZ was compared against several high-performing baselines, including STGCN, GraphWaveNet, ASTGCN, and DGCRN.
Performance During Disruptions
In the Richmond dataset, GCN-RWZ demonstrated superior performance during work zone events. For a 45-minute forecast horizon, it achieved a Mean Absolute Error (MAE) of 1.16, significantly lower than the 1.35 achieved by the next best model, DGCRN. Qualitative analysis via RMSE heatmaps showed a consistent error reduction of approximately 25% during complex rush-hour disruptions.
Long-Term Stability
The advantage of GCN-RWZ became more pronounced as the prediction horizon increased. While baseline models saw rapid performance degradation under the "noise" of construction, GCN-RWZ maintained stability up to a 180-minute horizon. This suggests that the "Speed Wave" approach successfully isolates the underlying traffic state from the transient disruptions caused by roadwork.
Conclusion
GCN-RWZ demonstrates that integrating external modalities like construction schedules into deep learning architectures is vital for robust traffic forecasting. Future work will explore expanding this fusion framework to include other variables such as weather conditions, lane closure specifics, and accident reports to develop a comprehensive "all-weather" traffic forecasting system.
Summary updated