Predicting Urban Traffic Volumes Using Support Vector Regression

Improving the accuracy of traffic flow predictions is a task of significant interest due to its diverse benefits and applications in transportation engineering. In this study, the authors develop a series of Support Vector Regression (SVR) models and assess the effect of different kernel functions to address the following needs: determine the accuracy of Split Cycle Offset Optimization Technique (SCOOT) flow data; improve the precision of SCOOT flows so as to allow agencies to extend their use; quantify the predictive power of models that do not include SCOOT data and can be applied to any travel link; and gain an insight into the strengths, weaknesses of SVR compared to other widely used methods including Generalized Linear Models (GLZ), k-Nearest-Neighbors (KNN), and Dynamic Recurrent Neural Networks (DRNN). The results of a preliminary analysis showed that the inherent Mean Absolute Percentage Error (MAPE) of SCOOT flows is 14.4%, compared to the more reliable Automatic Traffic Counter (ATC) flows. A nonlinear Gaussian kernel-based SVR model improved the accuracy of SCOOT flows by 53% (MAPE). The results revealed the nonlinearity in the relationship between ATC flows and the independent variables that can be more effectively captured by nonlinear models and kernels. All models and kernels examined exhibit better performance in the case of high flows compared to medium and low traffic volumes. From the comparison of all “non-SCOOT-based” models, it was found that the nonlinear DRNN models outperform the GLZ and the KNN models, but not the SVR models which result in a MAPE of 12.5%.

• Supplemental Notes:
  • This paper was sponsored by TRB committee AHB45 Traffic Flow Theory and Characteristics. Alternate title: Predicting Urban Traffic Volumes Using Support Vector Regression (SVR).
• Corporate Authors:

  Transportation Research Board

  500 Fifth Street, NW
  Washington, DC  United States  20001
• Authors:
  • Tsapakis, Ioannis
  • Haworth, James
  • Wang, Jiaqiu
• Conference:
  • Transportation Research Board 94th Annual Meeting
  • Location: Washington DC, United States
  • Date: 2015-1-11 to 2015-1-15
• Date: 2015

Language

• English

Media Info

• Media Type: Digital/other
• Features: Figures; References; Tables;
• Pagination: 20p
• Monograph Title: TRB 94th Annual Meeting Compendium of Papers

Subject/Index Terms

• TRT Terms: Accuracy; Mathematical prediction; Traffic flow; Traffic volume; Urban areas
• Identifier Terms: SCOOT (Computer program)
• Uncontrolled Terms: Support vector regression
• Subject Areas: Highways; Operations and Traffic Management; Planning and Forecasting; I71: Traffic Theory;

Filing Info

• Accession Number: 01559869
• Record Type: Publication
• Report/Paper Numbers: 15-4725
• Files: TRIS, TRB, ATRI
• Created Date: Apr 6 2015 7:17PM