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

Traffic congestion has significant economic, environmental, and socialramifications. Intersection traffic flow dynamics are influenced by numerousfactors. While microscopic traffic simulators are valuable tools, they arecomputationally intensive and challenging to calibrate. Moreover, existingmachine-learning approaches struggle to provide lane-specific waveforms oradapt to intersection topology and traffic patterns. In this study, we proposetwo efficient and accurate "Digital Twin" models for intersections, leveragingGraph Attention Neural Networks (GAT). These attentional graph auto-encoderdigital twins capture temporal, spatial, and contextual aspects of trafficwithin intersections, incorporating various influential factors such ashigh-resolution loop detector waveforms, signal state records, drivingbehaviors, and turning-movement counts. Trained on diverse counterfactualscenarios across multiple intersections, our models generalize well, enablingthe estimation of detailed traffic waveforms for any intersection approach andexit lanes. Multi-scale error metrics demonstrate that our models performcomparably to microsimulations. The primary application of our study lies intraffic signal optimization, a pivotal area in transportation systems research.These lightweight digital twins can seamlessly integrate into corridor andnetwork signal timing optimization frameworks. Furthermore, our study'sapplications extend to lane reconfiguration, driving behavior analysis, andfacilitating informed decisions regarding intersection safety and efficiencyenhancements. A promising avenue for future research involves extending thisapproach to urban freeway corridors and integrating it with measures ofeffectiveness metrics.