Related Experiment Videos
A temporal causal deep learning for real-time traffic risk prediction on freeway merging areas
Ziqiu Sun1, Weiwei Qi1, Hanlin Jiang1
1School of Civil Engineering and Transportation, South China University of Technology, Guangzhou 510641 China.
Accident; Analysis and Prevention
|July 6, 2026
Summary
This study introduces a temporal causal deep learning framework for predicting freeway merging risks. The new model improves accuracy and provides interpretable insights into conflict risk propagation for safer driving.
Area of Science:
- Traffic Safety Engineering
- Artificial Intelligence in Transportation
- Deep Learning for Predictive Modeling
Background:
- Current aggregate safety indicators for conflict risk lack spatiotemporal resolution and causal interpretability.
- Existing deep learning models struggle with understanding the causal relationships in traffic conflict prediction.
- Freeway merging areas are critical zones requiring enhanced risk assessment for proactive safety interventions.
Purpose of the Study:
- To develop a temporal causal deep learning framework for accurate and interpretable risk prediction in freeway merging areas.
- To address the limitations of low spatiotemporal resolution and lack of causal interpretability in current methods.
- To enhance proactive safety management strategies through a better understanding of conflict risk dynamics.
Main Methods:
- Utilized high-resolution vehicle trajectory data from unmanned aerial vehicles to create multivariate time series features.
- Developed a TC-iTransformer model, integrating temporal causal convolutions and causal attention, building on the iTransformer architecture.
- Incorporated Regression Relevance Propagation (RRP) for inferring cross-lag causal structures and temporal interaction patterns.
Main Results:
- The TC-iTransformer model achieved superior prediction accuracy with MAE of 0.602 and RMSE of 0.910.
- Demonstrated significant improvements over a baseline model, reducing MAE by 17.3% and RMSE by 14.1%.
- Interpretation results revealed key causal pathways and lag structures in risk index evolution, offering time-resolved conflict propagation insights.
Conclusions:
- The proposed temporal causal deep learning framework effectively predicts driving conflict risk with high accuracy and interpretability.
- The model provides valuable insights into the temporal dynamics and causal factors influencing freeway merging safety.
- Findings support the development of targeted interventions for identifying risky behaviors and improving freeway safety management.
