Comparison of machine learning-based models for predicting non-motorized vehicle speeds within signal-controlled intersections

Urban road intersections are the high incidence of traffic conflicts and accidents, and non-motorized vehicles have become an important influence on the efficiency and safety of intersection traffic operation due to their dynamic and random nature. As an important parameter of non-motorized traffic flow operation, accurate prediction of vehicle speed is necessary. This study proposes a comparative research method based on machine learning for the prediction model of non-motorized vehicle speed in signal-controlled intersections. Firstly, the non-motorized vehicle operation data of two intersections in Guangzhou were obtained through video collection; secondly, the core influencing factors were screened out by double validation of Random Forest and Lasso regression; subsequently, four prediction models, namely, Decision Tree, Random Forest, GBDT and XGBoost, were constructed. The experimental results show that the XGBoost model performs optimally on the test set, with a prediction accuracy of 78.3%, RMSE(1.08) and MAE(0.79), which are 2.7% and 9.2% lower than that of the decision tree, and the Nash-Sutcliffe efficiency coefficients are significantly better than the other models. The result provides technical support for intersection signal timing optimization and safety control of networked vehicles.