Network-Level Vehicle Delay Estimation at Heterogeneous Signalized Intersections

Accurate vehicle delay estimation is crucial for assessing signalized-intersection performance and guiding traffic management. Traditional machine-learning models often assume identical distributions between training and testing data. This assumption is rarely met across intersections due to differences in geometry, signal timing, and driver behavior, leading to poor generalization. To address this, this study proposes a domain adaptation (DA) framework that leverages a small labeled subset from the target intersection to improve delay estimation across diverse intersections. This study introduces a novel DA model, gradient boosting with balanced weighting (GBBW), which reweights source-domain data based on similarity to the target domain, enhancing adaptability. Using data from 57 heterogeneous intersections in Pima County, Arizona, we demonstrate that GBBW outperforms eight state-of-the-art machine-learning regression models and seven instance-based DA methods, providing more accurate and robust delay estimates. This framework improves the transferability of machine-learning models, supporting more reliable traffic-signal optimization, congestion management, and performance-based planning in real-world transportation systems.