Using Smart Work Zone Trailer Data to Evaluate and Predict Lane Closure Impacts with a Consideration of Work Intensity

Project Details









Smart Work Zone Deployment Initiative

Principal Investigator
Natalia Ruiz-Juri

Director, Network Modeling Center, Center for Transportation Research, The University of Texas at Austin

About the research

This project describes the implementation of machine learning (ML) models to the prediction of work-zone traffic impacts including local speed and traffic volume changes and corridor-level travel time increases. It also summarizes efforts to refine an existing tool that estimates work-zone-related delays and costs by providing consistent estimates of typical travel times that consider variations across days of the week and months of the year.

All of the models described in the report were estimated/trained and tested using data collected on I-35 through Austin, Texas, on a 20.4-mile section on which smart work-zone trailers (SWZTs) were placed. Predictive models combined SWZT point speed and volume data with INRIX segment-level speed data. The researchers implemented artificial neural networks (ANNs) to forecast speed and volume changes for planned closures.

Speed forecasting models performed well on average (root mean square error [RMSE] of 10.19 mph) but tended to underestimate speed reductions when the closures were significant. The latter was likely a result of having a small fraction of time steps exhibiting significant speed reductions in the dataset, which consisted mostly of nighttime closures.

Models used to forecast changes in traffic volumes had an average error (RMSE) of 57 vehicles per hour per lane (vphpl), which was comparable to that of linear regression models. Further training with a more balanced dataset that includes daytime and nighttime closures is required to support a broader set of applications.

The researchers also analyzed the performance of three short-term travel-time prediction (STTTP) methods, trained as part of a separate effort during work zones. The trained models, which included a time series approach and two types of ANNs, were very successful on average, outperforming traditional approaches by up to 50 percent during the peak period. While model performance was not as impressive for predicting travel times when work zones were present, preliminary results were promising with ML models consistently outperforming the traditional approaches.

Further model refinements to explicitly consider the presence of work zones and their characteristics are expected to improve model predictions. The efforts described in this project illustrate the potential value of emerging data sources and modeling techniques to support work-zone planning and management.