Citation

Li H, Cao Q, Song X, Hu H. J. Transp. Eng. A: Systems 2023; 149(7): e04023049.

Copyright

(Copyright © 2023, American Society of Civil Engineers)

DOI

10.1061/JTEPBS.TEENG-7635

PMID

unavailable

Abstract

Accurate incident impact prediction is one of the crucial tasks in road safety management that contributes to mitigating the resultant congestion and accelerating emergency rescue. However, the impact of traffic incidents is affected by numerous and redundant factors, and their incompleteness and scalability make the task of impact prediction particularly challenging. To address these issues, a novel multimodule dynamic prediction framework, named information enhancement cascade forest (IECF), is proposed for traffic incident impact prediction in this paper. In the IECF model, XGBoost is adopted to screen the original incident factors to reduce the information redundancy. Additionally, a novel enhancement generative adversarial network (EGAN) was custom designed to complete the missing factors by historical large-sample adversarial training. Then a multigrained permutation cascade forest (GPCF) was constructed to predict the incident impact degree through integrated learning. Furthermore, sufficient comparative experiments were conducted, and the results show the proposed model has better performance on traffic incident impact dynamic prediction than the state-of-art methods, especially in partial-factors-available situations.

Language: en

Keywords

Cascade forest; Generative adversarial network; Incident impact prediction; Traffic incident