Predicting Freeway Incident Duration Using Machine Learning

Abstract

Traffic incident duration provides valuable information for traffic management officials and road users alike. Conventional mathematical models may not necessarily capture the complex interaction between the many variables affecting incident duration. This paper summarizes the application of five state-of-the-art machine learning (ML) models for predicting traffic incident duration. More than 110,000 incident records with over 52 variables were retrieved from Houston TranStar data archive. The attempted ML techniques include: regression decision tree, support vector machine (SVM), ensemble tree (bagged and boosted), Gaussian process regression (GPR), and artificial neural networks (ANN). These methods are known to effectively handle extensive and complex datasets. Towards achieving the best modeling accuracy, the parameters of each of these models were fine-tuned. The results showed that the SVM and GPR models outperformed other techniques in terms of the mean absolute error (MAE) with the best model scoring an MAE of 14.34 min. On the other hand, the simple regression tree was the worst overall model with an MAE of 16.74 min. In terms of training time, a considerable difference was found between two groups of models: regression decision tree, ensemble tree, and ANN on one hand and SVM and GPR on the other. The former required shorter training time (less than one hour each) whereas the latter had training times ranging between 5 to 34 hours per model.