Combining predictive schemes in short-term traffic forecasting


Type de document
Résumé / Abstract
The principal motivation for combining forecasts which can either be a class label (classification) or numerical (regression) has been to avoid the a priori choice of which forecasting method to use by attempting to aggregate (weighted averaging in regression and voting in classification) together all the information which each forecasting model embodies. In selecting the 'best' model (in some sense), the forecaster is often discarding useful independent evidence in those models which are rejected. Hence the methodology of combining forecasts is founded upon the axiom of maximal information usage. Short-term traffic prediction is an area where combination of two ore more predictions is a promising technique to directly improve the forecast accuracy. This approach may eventually help in specifying underlying processes more appropriately and thus build better individual models. Since the combination of predictors has, for the most part, implicitly assumed a stationary underlying process, attention has been focused on taking into account the effect of nonstationary of the traffic flow process. This article deals with a discussion about forecast combination methods potentially suitable for short-term prediction with their performance comparisons. The emphasis lies on the application to the short-term traffic flow prediction. It has been noted that the methods outlined in this paper improve the quality of the resulting forecast by reducing its error. At worst, the quality (in error reduction sense) of the combined forecast is comparable to that obtained if the 'best' model is chosen (in mean square sense).

puce  Accès à la notice sur le portail documentaire de l'IFSTTAR

  Liste complète des notices publiques de l'IFSTTAR