Development and Evaluation of Statistical and Machine-Learning Models for Queue-Length Estimation for Lane Closures in Freeway Work Zones

• Published in: Journal of construction engineering and management 2023-05, Vol.149 (5)
• Main Authors: Biswas, Pronab Kumar, Khan, Sakib Mahmud, Piratla, Kalyan, Chowdhury, Mashrur
• Format: Article
• Language: English
• Subjects: Accuracy; Configurations; Estimation; Heavy vehicles; Highway maintenance; Lane closure; Machine learning; Queues; Regression models; Rehabilitation; Statistical analysis; Statistical models; Technical Papers; Traffic flow
• ISSN: 0733-9364; 1943-7862
• DOI: 10.1061/JCEMD4.COENG-12648

AbstractFreeway maintenance and rehabilitation work usually require closing one or multiple lanes, interrupting traffic flows, and creating queues upstream of the work zone. Public agencies can use queue length as a criterion to determine the maximum duration of lane closures and necessary traffic diversions. Previous studies of estimating queue length due to work zone lane closures are data- and time-intensive. This study presents an efficient approach for estimating queue length estimation due to work zone lane closures by developing various statistical and machine-learning models. The inputs for these queue length estimation models were vehicle demand, lane closure duration, active work zone length, and heavy vehicle percentage. The extent of the queues caused by short-term work zones on freeways for 2-to-1 (one-lane closure on a two-lane freeway), 3-to-1 (one-lane closure on a three-lane freeway), and 3-to-2 (two-lane closure on a three-lane freeway) lane-closure configurations can be estimated with these models. The primary scientific contribution of this study is the applicability of the queue length estimation models in any freeway network with work zone configurations and geometric features such as those used for model development. This research evaluated the efficacy of both statistical and machine-learning models for estimating the queue length considering different work zone scenarios. The accuracy of the queue length estimation models was evaluated for a different network that the original models had not seen previously. Among the statistical models, the quantile regression model had the best accuracy based on mean absolute percentage error (MAPE) for the 2-to-1 lane-closure configuration (88%), and the multiple linear regression had the best accuracy for the 3-to-1 (76%) and 3-to-2 (72%) lane-closure configurations. Among the machine-learning models, the stacking regressor model had the best accuracy for 2-to-1 (95%), 3-to-1 (90%), and 3-to-2 (89%) lane-closure configurations. Based on the analysis, it was observed that machine-learning models performed better than the traditional statistical models in estimating queue lengths.