Introduction
Predicting Origin-Destination (O-D) traffic flows is a fundamental challenge in intelligent transportation systems. Accurate O-D matrices allow for optimized traffic management, congestion mitigation, and improved public transit planning. However, O-D flows cannot be measured directly; they must be inferred from observable link counts (e.g., from loop detectors). This inference is a classic "underdetermined" problem, where multiple trip patterns can produce the same observed link counts. Furthermore, traffic is highly dynamic, characterized by complex spatial-temporal correlations that traditional linear models struggle to capture.
This paper proposes a novel hybrid framework that combines Fusion Line Graph Convolutional Networks (FL-GCNs) with a Deviation-based Kalman Filter to achieve state-of-the-art accuracy in dynamic O-D prediction.
Methodology
The proposed architecture integrates deep learning's pattern recognition capabilities with classical state-estimation techniques. The framework consists of two primary components: the FL-GCN for spatial-temporal feature extraction and a Kalman Filter for error correction and uncertainty minimization.
Fusion Line Graph Convolutional Networks (FL-GCNs)
Standard Graph Convolutional Networks (GCNs) typically operate on nodes. However, in transportation, traffic data is naturally associated with road segments (links). To address this, the authors utilize a Line Graph Transformation, where the original road links are converted into nodes in a new graph structure. This allows the GCN to process the "flow" of information across the network topology more naturally.
The FL-GCN maps link flows to O-D pairs through a series of link-graph and node-graph convolutions. A critical feature of this model is "Historical Fusion," which integrates historical O-D patterns directly into the neural network, providing a strong prior for the prediction.
Deviation-based Kalman Filter
To stabilize the neural network's outputs, the framework incorporates a Kalman Filter that operates on deviations from historical averages. This ensures the model remains grounded in realistic traffic behavior while adjusting for real-time sensor data. A mixing parameter ($\lambda$) is used to balance the outputs of the FL-GCN and the Kalman Filter. Empirical results suggest that a weight of approximately 0.8 for the GNN component yields the most robust performance.
Experimental Evaluation
The model was evaluated using real-world data from the New Jersey Turnpike (NJTP), covering 26 interchanges and 50 links. Data from February to May 2013 was used for training, with June 2013 serving as the test set. Predictions were made at 15-minute intervals during peak morning hours.
Key Results
The Mixing FL-GCN consistently outperformed standalone Kalman Filters and standard GNN variants across 1-step, 2-step, and 3-step predictions (up to 45 minutes ahead):
- Accuracy: For 1-step prediction, the Mixing FL-GCN achieved a Mean Absolute Error (MAE) of 3.585, significantly lower than the standard Kalman Filter (4.019) and historical averages (3.942).
- Robustness: The model proved particularly effective in high-flow scenarios (O-D pairs with >100 vehicles), where it substantially reduced Root Mean Square Error (RMSE) compared to baseline methods.
Conclusion and Future Work
The study demonstrates that treating road segments as nodes in a line graph is essential for capturing the spatial dependencies of traffic flow. While the hybrid model shows significant promise, future work will focus on relaxing the assumption of constant vehicle speeds and extending the framework to complex urban networks where multiple route choices exist between O-D pairs. Additionally, incorporating external factors like weather and vehicle speed into the GNN could further enhance predictive accuracy.
Summary updated