Estimating lane utilization for variable approach lanes using explainable machine learning

Abstract

This study investigates the flow distribution patterns at intersection approaches with variable approach lanes (VALs) and proposes a new lane utilization adjustment factor specifically for VALs. Unmanned aerial vehicles (UAVs) were used to collect naturalistic data, including traffic flow, approach geometry, and signal control-specific information, at selected intersections in Hangzhou, China. Machine learning techniques were employed to develop accurate regression models for estimating lane utilization, separately for the two VAL statuses. Preliminary analyses indicate different VAL utilization patterns across sites, suggesting the presence of external factors, beyond the VAL control itself, influencing the VAL utilization. The machine learning regression models provide insights into these factors by ranking them based on the importance and the impact magnitude. From a practical standpoint, this study recommends the implementation of uniform lane guiding signs, lane geometry improvements, and driver education to enhance the operational efficiency of variable approach lanes.