Predicting traffic volume at urban intersections is critical to ensuring the stability and efficiency of the urban road network. To better capture spatio-temporal nonlinear correlation between traffic volume and contributing factors and improve the prediction performance, this study introduces a Dynamically Weighted LightGBM (Gradient Boosting Machine) framework (DW-LGBM) for forecasting hourly traffic volume at intersections. The proposed DW-LGBM model features three core improvements: a dynamic weight allocation component that captures nonlinear spatio-temporal dependency, a multi-dimensional feature engineering component that incorporates cyclical temporal trends, and a dual-stage noise suppression mechanism using Exponentially Weighted Moving Average (EWMA) and Kalman filtering to smooth the data. The proposed model is trained and tested with hourly traffic volume data collected from 209 urban intersections during 31 days in Chengdu, China. The results show that the predictions achieve superior performance metrics which surpass those of the baseline models (e.g., LSTM and XGBoost). The proposed architecture shows exceptional spatio-temporal adaptability for different urban intersections. However, it is found that all the models perform woeful in predicting traffic volume during peak hours due to the significant heterogeneity among intersections.
Bozhi et al. (Fri,) studied this question.