An application of regression tree methodology in freeway travel time estimation using speed as a proxy

Graduation date: 2009Accurate freeway travel time estimation is of increasing importance for the travelers’ information and route guidance system. A non-parametric statistical methodology known as regression trees is deployed in this research for dynamically and accurately estimating freeway travel times for the I5-I205 loop in the Portland Metro area of Oregon using speed as a proxy. In the absence of historical travel time data on PORTAL (Portland Oregon Regional Transportation Archive Listing), which is the source of data collection in this research, regression tree models are built to predict speeds first and the predicted speeds are in turn used to estimate travel times by mid-point algorithm. The regression tree models in this research are built based on historical data sets, including not only the traffic flow data but also the incident related data, weather data and time of day. This ensures the models will maintain stable prediction ability under both free flow conditions and non-free flow conditions on freeways. Model construction and validation are implemented in the statistical software package S-PLUS. A full regression tree model is constructed on one test data set including 227 daily test data sets randomly selected from the total of 342 daily test data sets collected in the entire year of 2005. To determine what kind of regression tree model should be selected to predict speed or estimate travel time for a certain day under dynamic conditions, a characterization approach is deployed and four characterization standards are setup to track the characteristics of both test data sets and validation data sets. Two experimental designs are constructed to evaluate and compare the performances of eleven regression tree models ─ the full regression tree model and\ud the ten characterization regression tree models. The results show that these eleven tree models possess the ability to accurately predict speeds or estimate travel times. In addition, meaningful results are obtained showing which of these eleven tree models are best to choose for dynamically estimating travel times for a future day