Travel time prediction in the presence of traffic incidents: A neural network approach.

摘要

Incident related traffic congestion leads to huge economic loss each year. To predict the traffic situation when an incident occurs and disseminate the information to the traveling public can alleviate the traffic congestion caused by the incident. However, relatively little research has been conducted on this topic due to the lack of reliable data sources, poor data fusion techniques, and modeling difficulties.; The current research develops an artificial neural network framework for travel time prediction in the presence of traffic incidents. Several techniques, including input channel screening, cross validation and genetic algorithm, are utilized to optimize the model. Based on the fact that different typologies of neural networks specialize in different portions of the input pattern space, two novel complementary networks are proposed: Weighted Average Complementary Network (WACN) and Best Performer Complementary Network (BPCN). The proposed complementary networks combine different neural network typologies together and capitalize on the benefits of each of them. Traffic incident, traffic condition, and weather condition data from Interstate Highway 66 East Bound in Northern Virginia are collected and fused. Experiments are conducted under two scenarios: with and without traffic condition information in the input vectors. The performances of three individual networks and the two proposed complementary networks are examined and compared.; The results demonstrate that it is possible to accurately predict the future travel time within a corridor in the presence of traffic incidents given sufficient amount of data. With exceptional learning ability, neural network approach is proven to be an effective tool. The developed networks deliver a good fit in most cases, indicating that they are successful models. The proposed complementary networks further reduce both the prediction errors and the variations of the prediction errors, suggesting more accurate and reliable predictions. It is observed that the complementary networks perform best when the integrated individual networks give overall comparable but locally different results. It is also found that incident related information roughly dictates the trend of the impact on traffic, while current traffic condition provides a dynamic environment where the incident occurs. Consequently, addition of current traffic condition information can further improve the prediction accuracy.