Balancing Confidence and Precision: A Framework for Real-Time Bus Arrival Time Prediction with Uncertainty Quantification

Abstract

Accurate and reliable real-time bus arrival time (BAT) predictions are crucial for improving passenger satisfaction and operational efficiency. Existing pointwise BAT prediction models have demonstrated their effectiveness in estimating single values close to the true arrival time. However, there is a lack of research on quantifying the uncertainties associated with these predictions, which is essential for better passenger planning and enhancing the credibility and reliability of bus operators. This paper introduces UncertBAT, a novel framework designed to address this gap. UncertBAT provides not only the predicted BAT but also an arrival time window with a high degree of confidence. The model incorporates conformalized quantile regression and a grouping calibration mechanism to address challenges posed by data skewness and variability, ensuring an optimal balance between prediction confidence and precision. Several experiments conducted in the study demonstrate the model’s effectiveness in BAT prediction.