SPONSOR: Oregon Department of Transportation, OTREC

BUDGET: $85,300

ABSTRACT: The accurate and reliable prediction of travel times in traffic networks still remains to be the foremost challenge in intelligent transportation systems (ITS). Although flow and density/occupancy can be regarded as variables that are capable of explaining the variation in travel time, there is sufficient evidence that they are only applicable in uncongested regimes where traffic data suggest that speeds are relatively constant and other factors have not changed. Under congested regimes, the departure from this relationship is severe and it is further exacerbated by other variables. These include several independent (both indicator and measurable) variables such as the day, time of the day, weather conditions, incidents, elapsed time since the incident, construction work zones, lane closures, and special events such as sports amongst others. Previous research on travel time prediction has relied heavily on techniques such as simulation, linear regression, and sophisticated neural network models. Simulation models require a lot of human effort and are not easily generalizable. Linear regression models are too simplistic and neural networks, while being accurate, are hard to train. In this research, we propose to build prediction models that belong to a new class of nonlinear regression, characterized as regression trees, and compare its prediction performance to linear regression. We intend to use ODOT data collected and archived at Portland State University for the Interstate 5-205 (I5-I205) loop in the Portland area to perform the experiments and develop the regression tree based models. We anticipate the proposed exploratory research to uncover insightful directions that would lay the foundation for performing further research in travel time prediction in complex interconnected traffic networks.