Predicting traffic delays caused by lane-changing behaviour on multilane highways using machine learning models

Lane-changing behaviour on multilane highways significantly influences traffic efficiency and safety. Each lane change generates disturbances that propagate through traffic, causing deceleration waves and increased travel time. This study quantifies and predicts the time delay experienced by target lane following vehicles (TLFVs) resulting from adjacent lane-changing manoeuvres under free and moderate traffic conditions. Using the HighD dataset, which contains high-resolution drone-based vehicle trajectory data from German highways, lane-change events were extracted to engineer features such as vehicle type, minimum gap, lane-change duration, and traffic density. Three predictive models; Linear Regression, Random Forest, and XGBoost Regression, were developed and compared using RMSE, MAE, and R2. Following systematic hyperparameter tuning using Repeated K-Fold Cross-Validation, results showed that Random Forest Regression achieved the best performance R2= 0.86), followed by XGBoost R2= 0.82), with both ensemble methods significantly outperforming the Linear Regression baseline. Feature importance analysis revealed that minimum gap and lane-change duration were the most influential predictors of delay, while interactions involving heavy vehicles produced the highest average delays. The proposed models demonstrate the robust capability of ensemble machine learning to capture nonlinear traffic interactions and provide accurate delay predictions. These findings can support intelligent transportation systems (ITS) and guide future work in real-time delay estimation and adaptive traffic control.