Video-Based Comparative Modelling of Bus Dwell Time Using Parametric and Machine Learning Methods

Abstract

This study models bus dwell time and compares different modelling techniques to identify key operational and passenger-related factors affecting it. Using manually collected stop-level data from İzmir, Türkiye, the analysis examined four main explanatory factors, namely, total boardings, in-vehicle crowding level, number of doors opened, and passenger complexity ratio. After data cleaning and outlier screening, the final modelling dataset included 381 observations. Parametric models, including Multiple Linear Regression (MLR), Elastic Net Regression (ENR), and Polynomial Regression (PR), were compared with non-parametric machine learning models, including k-Nearest Neighbors (KNN), Random Forest (RF), and XGBoost. Model performance was evaluated using R2, adjusted R2, Mean Squared Error, Mean Absolute Error, and Root Mean Squared Error. Among the parametric models, ENR provided the strongest performance, with an adjusted R2 of 71.4%. The ensemble-based models performed best overall, with XGBoost achieving the highest adjusted R2 of 77.5%, followed closely by RF. SHAP analysis confirmed that total boardings was the dominant predictor, followed by the number of doors opened, passenger complexity ratio, and high in-vehicle crowding. The TCQSM-style benchmark comparison showed that locally calibrated and data-driven models provided more accurate dwell-time estimates than externally parameterized service-time values. The findings show that video-based stop-level observations can support both interpretable and prediction-oriented bus dwell-time modelling.