Related Experiment Videos
Multisource Urban Sensing Data Fusion and Dynamic Causal Graph Modeling for Explainable Traffic State Prediction
Ran Zhu1,2, Yingxi Wu2, Xiaoya Wang2,3
1The Bartlett School of Architecture, University College London, London WC1E 6BT, UK.
Sensors (Basel, Switzerland)
|July 28, 2026
Summary
This study introduces a novel spatiotemporal causal graph learning framework for urban traffic congestion prediction using multisource sensing data. The method significantly improves traffic state prediction accuracy and early warning reliability in smart cities.
Area of Science:
- Artificial Intelligence
- Urban Planning
- Data Science
Background:
- Urban traffic congestion prediction is crucial for smart city sensing and intelligent traffic governance.
- Existing methods struggle to capture dynamic congestion propagation due to reliance on single-source data or static road topology.
- Multisource sensing information (traffic flow, trajectories, images, public transport, weather, events) is vital for comprehensive analysis.
Purpose of the Study:
- To propose a spatiotemporal causal graph learning framework for urban traffic state prediction, congestion identification, and explainable early warning.
- To effectively fuse and leverage multisource urban sensing data for enhanced traffic analysis.
- To improve the accuracy and reliability of traffic congestion prediction and early warning systems.
Main Methods:
- Developed a collaborative encoding module to fuse multisource urban sensing data, using a reliability-aware attention mechanism to handle low-quality or missing data.
- Implemented a dynamic spatiotemporal causal graph learning module to adaptively learn time-varying causal relationships from historical traffic states, road topology, and external disturbances.
- Employed a causality-explanation-driven congestion prediction module to jointly model spatial diffusion and temporal evolution, identifying congestion sources, paths, and factors.
Main Results:
- Achieved Mean Absolute Error (MAE) values of 3.21, 3.79, and 4.48 for 15-min, 30-min, and 60-min traffic state predictions, outperforming existing state-of-the-art methods.
- Demonstrated high performance in congestion identification and early warning across complex scenarios (peak-hour, rainy weather, traffic events) with F1 scores up to 0.927 and AUC values up to 0.966.
- Significantly reduced the false alarm rate (FAR) in early warning systems, indicating improved reliability.
Conclusions:
- The proposed multisource spatiotemporal causal graph learning framework effectively enhances traffic state prediction accuracy and congestion early warning reliability.
- The method provides an interpretable approach by identifying key congestion sources, propagation paths, and inducing factors.
- This framework offers a robust technical pathway for AI-driven intelligent traffic sensing and management in smart cities.
