Utilizing data mining techniques to predict expected freeway travel time from experienced travel time

Abstract

As the most important real-time traveler information, travel time can be either experienced or expected (i.e. to be experienced). When a vehicle completes a trip, the travel time refers to the experienced travel time. In contrast, when a vehicle starts its journey, the travel time is unknown but can be predicted, which is the expected travel time. Although the experienced travel time is termed as the real-time travel time, a traveler may encounter a somewhat different travel time (from expected travel time) due to the changing traffic conditions. Therefore, expected travel time needs to be predicted. In this study, the expected travel time was predicted from the experienced travel time using the data mining techniques such as k-nearest neighbor (k-NN), least squares regression boosting (LSBoost) and Kalman filter (KF) methods. After comparing the performances of KF to corresponding modeling techniques from both link and corridor perspectives, it is concluded that the KF method offers superior prediction accuracy in a link-based model. Moreover, the effect of different noise assumptions was examined and it is found that the steady noise computed from the full-dataset had the most accurate prediction. A data processing algorithm, which processed more than a hundred million records reliably and efficiently was also introduced.