Understanding spatial dependency structure in urban road traffic networks: methodology and applications in short term traffic prediction

This study considers three approaches to identify relevant links in a large-scale urban network, namely, bivariate linear method, multivariate linear method and nonlinear method. We propose pairwise Granger causality test as the bivariate linear method, vector auto-regressive Granger causality test as well as elastic net regularisation as the multivariate linear method and regression tree as the nonlinear method for selecting relevant predictor links for the target links in the network. This study adopts pairwise Granger causality test as the bivariate linear method to capture dependence of the traffic states of a target link to the traffic states of other links in a road network and thus identify the relevant predictor links of the target link. The Granger causality test detects the relevant predictor links that have a statistical causal effect to the target link and the magnitude of the dependence of the target link can be measured using a metric called Granger-causal strength. We propose a variable selection method that utilises this Granger causal strength to select the most significant predictor links among the relevant links of the target link. The efficiency of the proposed variable selection method is demonstrated in terms of dimensionality reduction and prediction accuracy. The urban road network of Brisbane in Australia is selected as a test bed and 5-minute interval traffic flow data are used to demonstrate the application of the proposed variable selection methods in building short-term traffic prediction models. The case studies show that the proposed methods are effective in detecting spatial dependence among road links in the network and selecting a parsimonious set of relevant predictor links for an individual target link or a whole road network. The parsimonious set of relevant predictor links can then be used as the input variables in short-term traffic prediction models, which have reduced computational complexity while ensuring higher prediction accuracy.

Media Info

  • Pagination: 1 file

Subject/Index Terms

Filing Info

  • Accession Number: 01741795
  • Record Type: Publication
  • Source Agency: ARRB Group Limited
  • Files: ITRD, ATRI
  • Created Date: Jun 2 2020 12:07PM