An Improved Bus Travel Time Prediction Using Multi-Model Ensemble Approach for Heterogeneous Traffic Conditions

An accurate and reliable arrival time prediction of buses to the next bus stops is a valuable tool for both passengers and operators. Existing studies have some limitations in bus travel time prediction. They focus little on three aspects such as heterogeneous traffic flow conditions, dwell time prediction and interpretation of explanatory variables. Consequently, we break down the prediction problem into sub-models for running time and dwell time prediction and incorporate a feature engineering framework that generates features related to the running bus, the prediction day, and immediate and historical time variations to capture heterogeneous traffic conditions. We propose a multi-model stacked generalisation ensemble model by leveraging the advantages of best-performing models in homogeneous conditions such as Extreme Gradient Boosting (XGBoost) and convolutional long short-term memory (ConvLSTM) models. It outperformed the state-of-the-art models by 11% in mean absolute error (MAE) on average. It can predict extreme conditions in bus journeys more accurately. When interpreting the feature importance, we found insights like driver behaviour and preceding travel time influence the prediction model, which paves the way for strategic management by authorities.