Dynamic Graph Attention Networks for Travel Time Distribution Prediction in Urban Arterial Roads

TL;DR

The paper tackles the challenge of estimating arterial travel-time distributions for both directions along urban corridors. It introduces FDGNN, a Fusion-based Dynamic Graph Neural Network that uses static and dynamic graphs with intermediate feature fusion and a sequential optimization scheme to predict travel times as a normal distribution, $tt_i^e$ and $tt_i^w$, from corridor-state inputs. The architecture combines a static imputation module $M_x$ with dynamic modules $M_{\mu}$ and $M_{\sigma}$ to learn mean and dispersion, validated on a 9-intersection corridor with extensive SUMO simulations, demonstrating robustness to cycle length, volume, and green-time variations. The approach offers scalable, real-time capable traffic optimization by leveraging few, readily available inputs and a compact model footprint, enabling enhanced travel-time reliability in smart city applications.

Abstract

Effective congestion management along signalized corridors is essential for improving productivity and reducing costs, with arterial travel time serving as a key performance metric. Traditional approaches, such as Coordinated Signal Timing and Adaptive Traffic Control Systems, often lack scalability and generalizability across diverse urban layouts. We propose Fusion-based Dynamic Graph Neural Networks (FDGNN), a structured framework for simultaneous modeling of travel time distributions in both directions along arterial corridors. FDGNN utilizes attentional graph convolution on dynamic, bidirectional graphs and integrates fusion techniques to capture evolving spatiotemporal traffic dynamics. The framework is trained on extensive hours of simulation data and utilizes GPU computation to ensure scalability. The results demonstrate that our framework can efficiently and accurately model travel time as a normal distribution on arterial roads leveraging a unique dynamic graph representation of corridor traffic states. This representation integrates sequential traffic signal timing plans, local driving behaviors, temporal turning movement counts, and ingress traffic volumes, even when aggregated over intervals as short as a single cycle length. The results demonstrate resilience to effective traffic variations, including cycle lengths, green time percentages, traffic density, and counterfactual routes. Results further confirm its stability under varying conditions at different intersections. This framework supports dynamic signal timing, enhances congestion management, and improves travel time reliability in real-world applications.

Figures (4)

• Figure 1: The inputs and outputs used in the modeling of an arbitrary urban corridor. Inflow loop detectors are positioned 500 meters upstream of intersections. The diagram clearly distinguishes between input variables (in orange), output variables (in blue), and output variables that are reused as inputs in other modules (in purple). Input variables include traffic volume over a certain interval from inflow waveform time series in three directions upstream of an intersection, signal timing parameters (e.g., cycle length, offset, maximum green duration), driving behavior parameters (e.g., speed, acceleration, space cushion, lane changing, etc.), turning movement counts, and the distance between consecutive inflow loop detector locations along the main route. Reused output variables include traffic volume of intervening road segments between intersections on the arterial thoroughfare. Output variables include the normal distribution of eastbound and westbound arterial travel times.
• Figure 2: Graphical representation of traffic state of an urban corridor. Important factors are attributed as features to nodes and edges to uniquely represent the traffic state of an arbitrary urban corridor as a bidirectional and acyclic (dynamic) graph with k nodes indexed by $\{J^1,..J^k\}$ and 2k edges. The edge features are attributed to the direction of the movement between intersections.
• Figure 3: Overview of the proposed framework. This diagram illustrates the architecture of the proposed FDGNN framework, which consists of three modules. The $M_x$ module processes static graph data with masked node features representing intervening traffic volumes, reconstructing the node features. These reconstructed features are then passed as inputs to the $M_{\mu}$ and $M_{\sigma}$ fusion-based modules. The framework's outputs are a discrete normal probability density function (PDF) of bidirectional arterial travel time. The modules are sequentially optimized and the final loss is computed with respect to the fitted travel time histograms obtained from traffic simulations on the target urban corridor.
• Figure 4: Visualizing the results of FDGNN. Comparison of actual (red) and predicted (green) curves by FDGNN for a single traffic scenario. The analysis includes the imputation of intervening inflow waveforms (right column) and the estimation of the normal Probability Density Function (PDF) of travel time in both eastbound and westbound arterial directions through the urban corridor.