Long-range Traffic Condition Forecast usingFeature Space Projection Method

摘要

We propose a long-term traffic information forecast method for nationwide traffic informationservices. This method can deal with various 'day factors' which traffic conditions depend on,e.g., days of the week, seasons, and vacation times. It also has advantages in both the size ofthe database and the amount of calculation needed for forecasting. The basic idea of themethod is that traffic data consists of the weighted sum of several 'feature bases' whichrepresent characteristic traffic patterns such as congestion in mornings or evenings. Featurebases are information which is independent of day factors; in contrast, the information to beforecasted according to day factors is weighting coefficients of the feature bases, whosedimension is lower than that of original traffic data. Evaluation results show that the methodefficiently improves the forecast accuracy, especially for complicated traffic conditions withheavy congestion.