Methods to Analyze and Predict Interstate Travel Time Reliability

About the report:

The Moving Ahead for Progress in the 21st Century Act (MAP-21) defined requirements for system reliability performance measures.  Under MAP-21, state departments of transportation are responsible for reporting travel time reliability and for setting targets and showing progress toward those targets. In order to know how to improve travel time reliability and what to expect from investments in transportation infrastructure, these agencies need a better understanding of the factors that affect travel time reliability and methods to predict future travel time reliability.  The purpose of this study was to quantify the factors influencing travel time reliability and investigate how to account for these factors insetting reliability targets and communicating progress.

To achieve these objectives, this study developed models to estimate quantiles (the 50th, 80th, and 90th) of travel time distributions to quantify the effects of travel time reliability impact factors and predict select reliability measures.  First, linear quantile mixed models (LQMMs)were built using both data maintained by the Virginia Department of Transportation (VDOT) and crowdsourced event data.  Model results using the crowdsourced data were unstable and difficult to interpret because of data quality issues such as unbalanced spatial density, duplicate reporting, and inconsistent event classification because of individual observer bias.  The results using VDOT-maintained data were more reliable and interpretable.  Those models showed that frequencies of non-recurrent events, such as incidents and weather, were correlated with higher travel time percentiles.  The LQMM was compared with the trend line approach, a common prediction method used in practice, and the results showed that LQMMs significantly improved the accuracy of predictions over the trendline approach based on mean absolute percent error.  Generalized random forest (GRF) models were also tested as an alternative prediction method.  GRF models improved the prediction accuracy over LQMMs for the 50th and 80th percentiles, but the accuracy was slightly worse for the 90th percentile.  In addition, the GRF models could also reflect the impact of variables that were removed from LQMMs because of insignificance, such as the presence of safety service patrols.

Before-after studies were conducted to illustrate the application of LQMMs and GRF models.  LQMMs captured the changes in the 90th percentile travel times better, and GRF models captured the changes of level of travel time reliability better in most cases.  GRF models were more sensitive to the reliability changes caused by non-recurrent events, such as incidents or work zones, and could reflect the impact of variables that were removed from LQMMs because of insignificance. 

In addition, further research is recommended to extend the GRF models to meet the requirements of MAP-21 federal target setting.

Authors