Urban Traffic Flow Prediction Using a Spatio-Temporal Random Effects Model

Feng Chen, Chang-Tien Lu

Abstract

Traffic prediction is critical for the success of intelligent transportation systems (ITS). However, most spatio-temporal models suffer from high mathematical complexity and low tune-up flexibility. This article presents a novel spatio-temporal random effects (STRE) model that has a reduced computational complexity due to mathematical dimension reduction, with additional tune-up flexibility provided by a basis function capable of taking traffic patterns into account. Bellevue, WA, was selected as the model test site due to its widespread deployment of loop detectors. Data collected during the 2 weeks of July 2007 from 105 detectors in the downtown area were used in the modeling process and traffic volumes predicted for 14 detectors for the entire month of July 2008. The results show that the STRE model not only effectively predicts traffic volume but also outperforms three well-established volume prediction models, the enhanced versions of autoregressive moving average (ARMA) and spatiotemporal ARMA, and artificial neural network. Even without further model tuning, all the experimental links produced mean absolute percentage errors between 8% and 16% except for three atypical locations. Based on lessons learned, recommendations are provided for future applications and tune-up of the proposed STRE model.