1. Home
  2. TRID
  3. View Record
https://www.nationalacademies.org/trb/events

Self-learning flow status forecast for the Ring Road I

Itseoppiva sujuvuusennuste Keha I:lle

The purpose of the study was to create a self-learning short-term prediction model for the flow status of the vehicles entering the road sections within the next 15 minutes. The purpose was to develop a model capable of learning from the traffic situations that it observes and of adjusting the forecasts according to them without saving all the measurement data into databases. The targets of the study were fulfilled, as the result was a model based on self-organising maps and clustering that could predict the flow status for the road sections. The structure made it possible for the model to learn from the traffic situations without a need to save all the measurements into databases. The classification and the updating of the output probability classes made it possible. When developing the model, it became obvious that a simple median could not filter all the distortions caused by outliers in the travel time observations. A simple method turned out to be efficient in the online preprocessing of the data: travel time median data was filtered according to the number of observations and how much the current travel time median relatively differed from the previous accepted median value. According to an online trial, the proportion of the forecasts that could not be made because the cluster of similar traffic situations was empty decreased in time due to the self-learning principle, just as expected. The decrement was for all road section approximately 0.1 percent per day. The model could have predicted the flow status better if information on flow rates had been available in addition to travel times. Now, the flow rate information was available but with such a long delay (often up to 20 minutes) that the delay cancelled out the benefits of the additional information. The operational principle of this model can be transferred easily to other locations. The traffic monitoring system of the location dictates the input candidates. The area from which the input is collected for a certain sub-model is dependent on the location and on the characteristics of traffic. Input candidates for the prediction model can be predefined according to an expert opinion but the final input variables should be selected according to the procedure explained in the report. The project has been granted European Community financial support in the field of Trans-European Networks - Transport. This report may be found at http://alk.tiehallinto.fi/julkaisut/pdf/3200910-vitseoppiva_sujuvuusennuste.pdf

Language

  • Finnish

Media Info

Subject/Index Terms

Filing Info

  • Accession Number: 01011830
  • Record Type: Publication
  • Source Agency: TRL
  • ISBN: 951-803-400-1
  • Files: ITRD
  • Created Date: Dec 20 2005 11:24AM