Dynamic driving risk quantification and Informer-based prediction for heavy vehicles during right turns at...

Area of Science:

• Traffic Safety Engineering
• Artificial Intelligence in Transportation
• Deep Learning for Predictive Analytics

Background:

• Right-turn crashes involving heavy vehicles at urban intersections pose significant safety challenges.
• Existing risk assessments often rely on static factors, neglecting dynamic risk prediction in high-risk scenarios.
• There is a need for dynamic risk quantification and effective prediction models for heavy vehicle driving behavior.

Purpose of the Study:

• To develop a dynamic risk quantification method for heavy vehicle right-turn scenarios at urban intersections.
• To introduce and evaluate an advanced deep learning model, the Informer, for predicting driving risks.
• To compare the Informer model's performance against other deep learning architectures for risk prediction.

Main Methods:

• A dynamic risk quantification approach using an entropy weighting method with sliding time windows was employed.
• The Informer deep learning model was utilized for predicting driving risk, with defined observation, interval, and prediction time windows.
• Model performance was validated by comparing the Informer against Long Short-Term Memory (LSTM), CNN-Bi-Attention-LSTM, and Transformer models.

Main Results:

• The Informer model demonstrated superior performance with an 82.69% prediction precision under specific time window configurations (10s observation, 1s interval, 5s prediction).
• The Informer model achieved a highly efficient inference time of 0.014 ms per sample.
• Significant improvements in accuracy and efficiency for long-term risk prediction were observed compared to LSTM, CNN-Bi-Attention-LSTM, and Transformer models.

Conclusions:

• The proposed method offers an effective approach for dynamic driving risk prediction for heavy vehicles at urban intersections.
• This study addresses the limitations of traditional static risk analysis in traffic safety.
• The findings provide valuable technical support for developing active safety management systems for urban traffic.