Learning Systems for Predicting Experiential Travel Times in the Presence of Incidents: Insights and Lessons Learned

Abstract

Implementation of the intelligent transportation systems program has resulted in a wealth of readily available data describing key transportation system parameters for many of the nation's roadways. However, the development of systems that can process the data to support effective decision making aimed at improving travel conditions has lagged behind the development of monitoring and data recording devices. A study investigated the feasibility of developing an artificial neural network (ANN) model to predict experiential travel time under transient traffic conditions, including accidents, by using readily available data from roadway sensors in real time. Computational experiments were performed to investigate the impact of several factors on the quality of the ANN's predictions: ( a) the time lag of the input data, ( b) the input variables, and ( c) the temporal resolution of the input variables. With the correct input data set and network parameters, the ANNs were able to predict reasonable values for experiential travel time. The study shows that the length of the time lag did not have a statistically significant effect on ANN performance, that speed appears to be the most influential input variable, and that varying the resolution of the input data produced similar errors for both higher and lower resolutions.