Inter-Urban Short-Term Traffic Congestion Prediction

This thesis compares the performances of traffic congestion prediction models. The prediction horizon of the models included is short-term, covering 5, 10, 15, 20, 25, or 30 minutes. The methods that were used to build the models are: the naïve ("do nothing") method, multi-linear regression (MLR), auto regression moving average (ARMA) time series analysis, multi-layer feed-forward (MLF) artificial neural networks (ANNs), radial basic function (RBF) ANNs, Elman (state-space) ANNs, self-organizing map (SOM) ANNs, and fuzzy logic (FL). The author tested these models using two clover-leaf junctions in the Netherlands in the month of May 2001. Data included mean vehicle speed, standard deviation of vehicle speed, and intensities per vehicle length class. The developed models were calibrated and cross-validated. The results showed that the SOM models, FL models, and ARMA models are equally as good as or outperformed by the naïve models, thus these models do not have use as congestion prediction tools. The supervised ANN models showed the best results and are the best choice for congestion prediction. Appendices reproduce details of the data collected.

  • Availability:
  • Corporate Authors:

    Netherlands TRAIL Research School

    P.O. Box 5017
    Delft,   Netherlands  2600 GA
  • Authors:
    • Huisken, Giovanni
  • Publication Date: 2006


  • English

Media Info

  • Media Type: Print
  • Features: Appendices; Figures; Photos; References; Tables;
  • Pagination: 288p

Subject/Index Terms

Filing Info

  • Accession Number: 01050062
  • Record Type: Publication
  • ISBN: 9036524415
  • Report/Paper Numbers: Trail Thesis Series T2006/8
  • Files: TRIS
  • Created Date: May 22 2007 11:49AM