THE USE OF NEURAL NETWORKS AND TIME SERIES MODELS FOR SHORT TERM TRAFFIC FORECASTING: A COMPARATIVE STUDY

Short term forecasts of traffic flow at a site may need to be based on inputs from several other sites and times. Neural networks can be readily used for this purpose, because they can be trained to describe the output levels from the input data, without any need to specify in advance what form the relationships should take. Conventional time series analysis techniques can also be used, but require the form of model to be pre-specified. The paper reports results of fitting a range of models to data for a number of links from the "instrumented city" of Leicester, using SCOOT data for 5-minute intervals extracted via the link to Nottingham University. Two test situations were investigated; for a site and its adjoining links, and for a chain of links. The statistical models used were ordinary linear regression (between the flow at a site and flow at adjoining sites in the previous time period); discounted linear regression (in which flows in the previous "n" time periods were included); classical Box-Jenkins techniques (auto-regressive integrated moving average, or ARIMA); and transfer functions (in which the flow is modelled as a time function of the flow at the prediction site and its neighbour). The treatment of missing values and the goodness of fit of the models are described; and it was concluded that, of these statistical techniques, ARIMA models generally gave the better performance. The neural network models used are briefly summarised. Using historic data alone for all the sites showed a substantial under-prediction in the peak, but otherwise fairly good adherence to the original flow profile; but the fit overall was substantially improved by including current data for the adjoining sites as well. (This result is not as surprising as it might seem at first sight, because, in an urban area, some traffic may well flow from one site to another within the same 5-minute intervals). Overall, the conclusion was reached that differences in performance were slight and that other considerations become the determining factor when choosing a forecasting method. Whilst ARIMA models have the advantage of a more explicit structure, neural network models can be easier to develop and incorporate into local code. For the covering abstract of the seminar see IRRD 862693.

• Availability:
  • Find a library where document is available. Order URL: http://worldcat.org/isbn/0860502546
• Corporate Authors:

  PTRC EDUCATION RESEARCH SERVICES LTD

  GLENTHORNE HOUSE
  HAMMERSMITH GROVE, LONDON  United Kingdom  W6 OLG
• Authors:
  • CLARK, S D
  • Dougherty, M S
  • KIRBY, H R
• Conference:
  • TRANSPORTATION PLANNING METHODS. PROCEEDINGS OF SEMINAR D HELD AT THE PTRC EUROPEAN TRANSPORT, HIGHWAYS AND PLANNING 21ST SUMMER ANNUAL MEETING (SEPTEMBER 13-17, 1993), UMIST. VOLUME P363
• Publication Date: 1993

Language

• English

Media Info

• Features: References;
• Pagination: p. 151-62

Subject/Index Terms

• TRT Terms: Artificial intelligence; Computer programs; Conferences; Forecasting; Traffic congestion
• ITRD Terms: 8525: Conference; 632: Congestion (traffic); 8626: Expert system; 132: Forecast; 8645: Software
• Subject Areas: Data and Information Technology;

Filing Info

• Accession Number: 00647898
• Record Type: Publication
• Source Agency: Transport Research Laboratory
• ISBN: 0-86050-254-6
• Files: ITRD, ATRI
• Created Date: Jun 17 1994 12:00AM