Graph Attention Network for Lane-Wise and Topology-Invariant Intersection Traffic Simulation

TL;DR

This paper tackles the need for fast, fine-grained modeling of intersection traffic that generalizes across topology. It introduces two Graph Attention Network-based Digital Twins, $G_{ext}$ and $G_{inf}$, to reconstruct exit and inflow lane-waveforms using stop-bar and upstream detectors, respectively. Trained on roughly $4\times 10^5$ hours of SUMO-based simulations plus real ATSPM data, the models achieve accuracy comparable to microsimulations while enabling rapid, topology-agnostic waveform estimation for signal timing optimization. The approach supports corridor-level planning via O-D matrix generation and integrates with existing optimization frameworks, with future work aimed at urban freeway extensions and richer MOEs.

Abstract

Traffic congestion has significant economic, environmental, and social ramifications. Intersection traffic flow dynamics are influenced by numerous factors. While microscopic traffic simulators are valuable tools, they are computationally intensive and challenging to calibrate. Moreover, existing machine-learning approaches struggle to provide lane-specific waveforms or adapt to intersection topology and traffic patterns. In this study, we propose two efficient and accurate "Digital Twin" models for intersections, leveraging Graph Attention Neural Networks (GAT). These attentional graph auto-encoder digital twins capture temporal, spatial, and contextual aspects of traffic within intersections, incorporating various influential factors such as high-resolution loop detector waveforms, signal state records, driving behaviors, and turning-movement counts. Trained on diverse counterfactual scenarios across multiple intersections, our models generalize well, enabling the estimation of detailed traffic waveforms for any intersection approach and exit lanes. Multi-scale error metrics demonstrate that our models perform comparably to microsimulations. The primary application of our study lies in traffic signal optimization, a pivotal area in transportation systems research. These lightweight digital twins can seamlessly integrate into corridor and network signal timing optimization frameworks. Furthermore, our study's applications extend to lane reconfiguration, driving behavior analysis, and facilitating informed decisions regarding intersection safety and efficiency enhancements. A promising avenue for future research involves extending this approach to urban freeway corridors and integrating it with measures of effectiveness metrics.

Figures (5)

• Figure 1: The physical location of Stop-bar and virtual location of Exit and Inflow loop detectors in the simulation of an intersection. We simulate the ATSPM time series waveform within 8-phase standard NEMA phasing intersections. We train two digital twins that can estimate downstream exit waveforms of every outflow lane of all directions ($G_{ext}$) or upstream inflow waveforms of every inflow lane in all directions ($G_{inf}$) simultaneously for an intersection with arbitrary topology and characteristics.
• Figure 2: Overview of a proposed digital twin as applied to a single graph of traffic simulation for standard NEMA phasing intersection. The right subplot shows architecture of $\mathbf{G_{ext}}$ model. Masked node features $X$ are fed into a self-attention module for temporal encoding and then into a single GAT layer for spatial encoding while incorporating all effective factors specific to this specific traffic scenario. The decoder imputes missing (masked) time series waveforms for the exit loop detectors in the reconstructed feature matrix $\hat{X}$. The left subplot shows the architecture of $\mathbf{G_{inf}}$ model. Here the input of the model is multi-layered graph data and the imputed entities in the reconstructed feature matrix $\hat{X}$ are time series waveforms for the inflow loop detectors.
• Figure 3: Kernel density estimation of traffic flow exiting from different lane groups of an intersection. The exit time series waveform estimated by $G_{ext}$ for 1000 randomly selected samples from the test set is filtered for a specific intersection, grouped by associated lane groups (LG16: major westbound, LG25: major eastbound, LG47: minor southbound, and LG38: minor northbound) and reduced into two-dimensional embedding. The Gaussian distribution estimated for different lane groups represents four distinct dynamic behaviors of traffic flow within that intersection.
• Figure 4: Average impact of features on the magnitude of estimated exit waveform. We map the reduced dimension of the output of the model to its effective input features using a multivariate linear regression model and use it to explain our $G_{ext}$ digital twin by SHAP.
• Figure 5: Comparing actual (green-colored) and predicted (red-colored) Exit time series waveforms. For a single intersection, predicted time series waveforms reasonably match the actual ones. The same models can estimate lane-wise waveforms at all directions of different intersections. The plot shows Exit waveforms estimated at a 5-second-bucket resolution for a sample intersection (J3).