Abstract
The literature is rich in travel time prediction because of its importance in intelligent transportation systems. Despite the proliferation of advanced methodologies, modeling the uncertainty of traffic conditions is still a challenge, especially during congested situations. Travel time reliability associated with its time-dependent variation provides a way to measure the system performance and has received extensive attention in recent years. In practice, one of the measures for travel time reliability is the identification of prediction intervals (PIs). The PI measurement has many potential applications in the development of systems that aim at disseminating real-time traffic information to travelers. From a management point of view, the PIs forecast the unreliable traffic periods and enable the selection of proper strategies to avoid or release possible traffic congestion. The generalized autoregressive conditional heteroscedasticity (GARCH) model has proved to have the ability to model the uncertainties in the literature. However, the model has some drawbacks in traffic forecasting. To improve the quality of travel time reliability forecasting, this paper proposes two univariate volatility models and compares their performance in generating high-quantity PIs. Travel time data collected from automatic vehicle identification stations located along U.S. Highway 290 in Houston, Texas, are used to examine each model's performance in travel time reliability forecasting. Study results indicate that all three models give reasonable PIs that could be used to indicate the variability of future traffic conditions. The statistical analysis and forecasting results indicate that the proposed Glosten–Jagannathan–Runkle GARCH model outperforms the other two models in constructing better PIs.