A Framework for Travel Speed Prediction Inclusive of Service Area Dwell Times

Abstract

Accurate modeling of travel speeds is crucial for optimizing roadway management, yet traditional methods overlook a key factor the influence of vehicle dwell times in service areas. This oversight introduces bias into speed measurements, impairing their utility for fine-grained traffic monitoring. To address this problem, we propose an innovative framework that integrates machine learning prediction of service area dwell times into travel speed calculation. We focus on a 9.3 km segment of a major highway in Fujian Province, China that includes the Qingyunshan service area. A Gradient Boosting Decision Tree model identifies vehicles entering the service area, while a Bayesian Backpropagation Neural Network predicts their dwell time. By adjusting the overall travel times using these predicted dwell times, our approach recovers normal driving behavior outside service areas. Experiments on electronic toll collection data from over 17 million transactions validate the framework’s effectiveness. The corrected travel speeds better reflect typical highway conditions and enable more precise assessment of traffic state across multiple time horizons. This study highlights the vital role of service area dwell time in travel speed modeling. Our solution provides a promising direction to enhance the fidelity of current prediction practices.

Keywords

• Travel speed
• machine learning
• service area
• dwell time
• traffic state