An Adaptive System for Predicting Freeway Travel Times

Abstract

This paper presents an adaptive system that embeds a Bayesian inference-based dynamic model (BDM) for predicting real-time travel time on a freeway corridor. Bayesian forecasting is a learning process that revises sequentially the state of a priori knowledge of travel time based on newly available information. The prediction result is a posterior travel time distribution that can be employed to generate a single-value (typically but not necessarily the mean) travel time as well as a con¯dence interval representing the uncertainty of travel time prediction. To better track travel time°uctuations during nonrecurrent congestion due to unforeseen events (e.g., incidents, accidents, or bad weather), the BDM is integrated into an adaptive control framework that can automatically learn and adjust the system evolution noise level. The experimental results based on real loop detector data of a freeway stretch in Northern Taiwan suggest that the proposed method is able to provide accurate and reliable travel time prediction under both recurrent and nonrecurrent tra±c conditions. availability and quality of real-time travel time information, this study aims at developing an e®ective and robust online travel time prediction method to provide accurate and reliable short-term travel time forecasts of freeway corridors under di®erent tra±c conditions.A rich body of literature has been devoted to the development of short-term travel time prediction approaches over the last three decades. van Hinsbergen et al. 3 conducted a broad literature review on the existing tra±c prediction methodologies. The explored techniques for short-term travel time prediction include three major categories: (i) parametric methods, such as linear regression, 4,5 time series models, 6 and Kalman¯ltering techniques, 7À10 Bayesian models, 11 (ii) nonparametric statistical methods, such as neural network models, 12À17 Support Vector Regression, 18,19 and (iii) hybrid integration methods. 20À23 Recently, to enhance model°exibility and adaptiveness to the underlying tra±c pattern changes, relevant studies have been conducted to update online model parameters based on real-time observations. Yang et al. 24 compared an adaptive recursive least-square (ARLS) method and a Kalman¯lter method for online tra±c speed prediction; both methods recursively adjust the coe±cients of o®line models to achieve better prediction accuracy. Their¯nding showed that the (nonzero) noise covariance matrix is important to a state-space model. Liu et al. 25 extended the state-space neural network (SSNN) by incorporating the extended Kalman¯ltering (EKF) algorithm. They compared their approach with two existing models, namely the Kalman¯ltering method by Chien and Kuchipudi 8 and the neural network model trained by LevenbergÀMarquardt method, and found that their model performance is better in terms of computation time and prediction accuracy. Along the same line, van Lint 15 proposed censored EKF and delayed EKF algorithms for online traveltime prediction by adaptively training t...