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

Yousefzadeh, Nooshin; Sengupta, Rahul; Karnati, Yashaswi; Rangarajan, Anand; Ranka, Sanjay · 2025 · OpenAlex-citations

DOI: 10.1109/tits.2025.3546810

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the computational inefficiency and calibration challenges of microscopic traffic simulators, which struggle to provide lane-specific traffic waveforms or adapt to varying intersection topologies. To mitigate these issues, the authors propose two "Digital Twin" models based on Graph Attention Neural Networks (GAT) for simulating intersection traffic flow. These models aim to estimate detailed traffic waveforms for any intersection approach and exit lane, serving as lightweight, reusable alternatives to computationally expensive microsimulations like SUMO or VISSIM. The primary application is traffic signal optimization, where rapid prediction of traffic impacts is crucial. The methodology involves constructing graph-based auto-encoders that leverage high-resolution loop detector data and signal timing information. The authors define two specific problems: Exit Waveform Estimation and Inflow Waveform Estimation. For exit estimation, the model ($G_{ext}$) uses a single-layer directed bipartite graph connecting stop-bar detectors to exit detectors. For inflow estimation, the model ($G_{inf}$) uses a multi-layer graph connecting stop-bar detectors to upstream inflow detectors located 500 meters away. Both models incorporate edge features representing signal timing plans, turning-movement counts, and driving behavior parameters. The architecture employs attention mechanisms to capture temporal dependencies within individual waveforms and spatial dependencies between lanes, allowing the models to handle arbitrary intersection topologies by using standardized adjacency matrices with dummy nodes for missing lanes. The models were trained on datasets generated from 40,000 hours of simulation across a 9-intersection corridor, utilizing both real-world controller logs and random traffic scenarios with varying signal timing plans. The experimental results demonstrate that the GAT-based digital twins perform comparably to microsimulations in accuracy while offering superior computational efficiency. The models successfully generalize to unseen intersections and traffic conditions, estimating exit and inflow waveforms within seconds. The self-attention module proved effective in temporal decoding, and the attention-based aggregation improved consistency with overall traffic trends. The models performed slightly better on random traffic datasets due to the wider range of variations compared to restricted real-world scenarios. The significance of this work lies in providing an efficient, topology-invariant tool for traffic engineering. By relying exclusively on stop-bar detectors, which are predominantly deployed in real-world settings, the digital twins ensure practical applicability. The ability to rapidly predict the impact of signal timing changes and driving behaviors on lane-wise platoons facilitates informed decisions for intersection safety and efficiency. This approach enables seamless integration into corridor and network signal timing optimization frameworks, overcoming the computational bottlenecks of traditional microsimulation tools. Future research directions include extending the approach to urban freeway corridors and integrating it with measures of effectiveness metrics.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• traffic density