Analysing functional connectivity and causal dependence in road traffic networks with Granger causality

Understanding the dependence structure of road segments is a key to building successful prediction models for traffic forecasting. In order to accurately predict the traffic state of a particular target link at a given time interval, the prediction model should incorporate traffic states of other links that are spatially and temporally correlated with the target link into the model structure. Given a potentially very large number of links in a network, however, identifying a subset of links whose traffic states are highly dependent is a challenging task. To tackle this problem, this paper proposes a statistical approach that uses Granger causality analysis to determine the causal dependence among time series data of link traffic flow measures and identify a group of links that are functionally connected to a given target link. The functional connectivity refers to the statistical dependence between two locations in terms of their observed traffic states, regardless of their structural connectivity, which refers to the static/physical connection or spatial adjacency determined by the underlying physical road network. As such, two distant links characterized by low structural connectivity can show high functional connectivity if traffic flow time-series from these two links show high statistical correlation or dependency. The Granger causality analysis has been widely applied to detect such a functional connectivity in various spatio-temporal systems. In this study, the Granger causality analysis is applied on the time-series of link measures (traffic volume or speed) collected from a road network in Brisbane, Australia in 2014 to discover the (causal) dependency structure of the links and understand dynamic changes in the dependence structure across different times of the day. The paper tests both bivariate and multivariate linear vector auto-regression (VAR) models to perform pair-wise and multivariate Granger causality analyses, respectively, and discusses the performance difference between these models. The study also discusses the impact of different choices of link measures (i.e., volume time-series vs. speed time- series) on the performance of identifying the causal structure and the capability of short term traffic prediction.


  • English

Media Info

  • Pagination: 19p
  • Monograph Title: 38th Australasian Transport Research Forum (ATRF 2016), Melbourne, 16th - 18th November 2016

Subject/Index Terms

Filing Info

  • Accession Number: 01627434
  • Record Type: Publication
  • Source Agency: ARRB
  • Files: ITRD, ATRI
  • Created Date: Feb 27 2017 10:08AM