TSGDiff: Traffic state generative diffusion model using multi-source information fusion

Huipeng Zhang

doi:10.1016/j.trc.2025.105081

TSGDiff: Traffic state generative diffusion model using multi-source information fusion

2025-05-01

0

PRE

AI

DOI：10.1016/j.trc.2025.105081

Original

Original request for help

Save

Abstract

En 中文

Accurate analysis and prediction of traffic states are fundamental and crucial for intelligent transportation systems, playing a significant role in enhancing the efficiency and safety of traffic systems. Advances in deep learning have promoted the development of traffic prediction. However, some traditional prediction methods primarily rely on historical traffic data to sequentially predict future traffic trends. While some also incorporate one or more influencing factors, such as weather and day of the week, as covariates, they often lack a unified fusion approach to model the impact of these covariates on future traffic states, and they are prone to error accumulation in long-term predictions. To address these challenges, we propose TSGDiff, a novel traffic state generative diffusion model using multi-source information fusion. The proposed method leverages an innovative diffusion model framework and integrates various sources of information, such as traffic data, weather, and weekdays, to enhance the accuracy of traffic state prediction. TSGDiff transforms historical spatiotemporal information and future environment information into feature representations using an attention-based spatiotemporal extraction module and a traffic semantic encoding module, respectively. These feature representations serve as guiding conditions for the diffusion model to generate traffic states. By incorporating the prediction horizon as an input parameter, TSGDiff directly generates future traffic states point-to-point, thereby avoiding error accumulation inherent in iterative prediction methods. To adapt the diffusion model to graph structure road network data, we introduce a Graph Attention U-Net (GAUNet) to capture the spatial correlations in traffic data. Experiments on real-world Beijing traffic datasets demonstrate that TSGDiff significantly outperforms baseline models for long-term predictions and performs comparably for short-term predictions.

Keywords:

Generative model

Traffic state generation

Multi-source information

Conditional diffusion model

Speed prediction

TSGDiff: Traffic state generative diffusion model using multi-source information fusion

Abstract

Journal

Organization