Predicting hourly traffic volume of urban signal intersections using Dynamically Weighted LightGBM

Area of Science:

• Transportation Engineering
• Data Science
• Urban Planning

Background:

• Accurate traffic volume prediction is essential for urban road network stability and efficiency.
• Existing models often struggle to capture complex spatio-temporal correlations in traffic data.
• Urban intersection traffic exhibits nonlinear dynamics influenced by various factors.

Purpose of the Study:

• To develop an advanced framework for forecasting hourly traffic volume at urban intersections.
• To improve prediction performance by capturing nonlinear spatio-temporal dependencies.
• To enhance the robustness of traffic volume prediction models.

Main Methods:

• Introduction of a Dynamically Weighted LightGBM (DW-LGBM) framework.
• Incorporation of a dynamic weight allocation component for nonlinear spatio-temporal dependency.
• Implementation of multi-dimensional feature engineering and a dual-stage noise suppression mechanism (EWMA and Kalman filtering).

Main Results:

• The DW-LGBM model demonstrated superior prediction performance compared to baseline models like LSTM and XGBoost.
• The proposed architecture exhibited exceptional spatio-temporal adaptability across different urban intersections.
• All tested models showed poor performance during peak hours due to significant intersection heterogeneity.

Conclusions:

• The DW-LGBM framework offers a promising approach for improving urban intersection traffic volume forecasting.
• Addressing the challenges posed by intersection heterogeneity is crucial for enhancing peak hour prediction accuracy.
• Further research is needed to refine models for highly variable traffic conditions.