Traffic congestion has been a major problem in metropolitan areas, which is caused by either insufficient roadway capacity or unforesceable incidents. In order to promote the efficiency of the existing roadway networks and mitigate the impact of traffic congestion, the development of a sound prediction model for travel times is desirable.

A comprehensive literature review about existing prediction models was conducted by investigating the advantages, disadvantages, and limitations of each model. Based on the features and properties of previous models, the base models including exponential smoothing model (ESM), moving average model (MAM), and Kalman filtering model (KFM) are developed to capture stochastic properties of traffic behavior for travel time prediction.

By incorporating KFM into ESM and MAM, three dynamic recursive prediction models including dynamic exponential smoothing model (DESM), improved dynamic exponential smoothing model (JDESM), and dynamic moving average model (DMAM) are developed, in which the time-varying weight parameters are optimized based on the most recent observation. Model evaluation has been conducted to analyze prediction accuracy under various traffic conditions (e.g., free-flow condition, recurrent and non-recurrent congested traffic conditions). Results show that the IDESM in general outperforms other models developed in this study in prediction accuracy and stability. In addition, the feature and logic of the IDESM lead to its high transferability and adaptability, which could enable the prediction model to perform well at multiple locations and deal with complicated traffic conditions.

Besides the proficient capability, the IDESM is easy to implement in the real world transportation network. Thus, the IDESM is proven an appealing approach for short-time travel time prediction under various traffic conditions. The application scope of the IDESM is identified, while the optimal prediction intervals are also suggested in this study.