Prediction of Traffic Incident Duration Using Clustering-Based Ensemble Learning Method

Abstract

Traffic incidents are a primary cause of traffic delays, which can cause severe economic losses. Effective traffic incident management requires integrating intelligent traffic systems, information dissemination, and the accurate prediction of incident duration. This study develops a clustering-based machine learning model to predict the incident duration. Unlike similar studies that train separate machine learning models for a fixed number of clusters, this study proposes an ensemble learning method based on multiple clustered individual models that can provide good and diverse prediction performance. The K-means clustering method is used in this study as a bootstrapping technique in the ensemble learning approach, with the individual models based on the artificial neural network model and random forest regression model. The models are tested using the incident data from Singapore, and the results show that the ensemble model outperforms both the traditional model with fixed clusters and the classical model without clustering. Additionally, this study attempted to determine the significance of different variables on traffic incident durations using the random forest feature importance function. The prediction of incident duration and the analysis of influence factors can contribute to several aspects of traffic management, such as improving traffic dissemination to mitigate traffic congestion caused by incidents.