Roadway traffic crash prediction using a state-space model based support vector regression approach.

Area of Science:

• Transportation Engineering
• Traffic Safety Analysis
• Predictive Modeling

Background:

• Conventional traffic crash analysis methods often overlook dynamic roadway system changes, leading to inaccurate predictions.
• Existing models struggle to account for evolving traffic flow and its impact on crash risk factors.
• There is a need for advanced predictive models that capture the temporal dynamics of roadway systems.

Purpose of the Study:

• To propose an innovative two-step method for enhanced traffic crash prediction.
• To integrate a state-space model (SSM) with a support vector regression (SVR) model for improved accuracy.
• To address the limitations of traditional methods by incorporating dynamic roadway system evolution.

Main Methods:

• A two-step approach was developed: first, an SSM to model roadway system dynamics and predict impact factors.
• Second, an SVR model used predicted impact factors for traffic crash prediction.
• The model was validated using a five-year dataset from 1152 roadway segments in Tennessee.

Main Results:

• The proposed model achieved an average prediction MAPE of 7.59%, MAE of 0.11, and RMSD of 0.32.
• It demonstrated superior prediction accuracy compared to standalone SVR and multivariate negative binomial (MVNB) models.
• Prediction accuracy improved by 4.360% (vs. SVR) and 6.445% (vs. MVNB) on average, with enhanced robustness and smoother predictions.

Conclusions:

• The integration of SSM for dynamic roadway system analysis significantly enhances traffic crash prediction accuracy.
• The proposed two-step method offers a more robust and accurate approach compared to conventional models.
• The model effectively handles data variations and the phenomenon of extra zeros, crucial for real-world traffic safety.