Traffic flow breakdown is the abrupt shift from operation at free-flow conditions to congested conditions and is typically the result of complex interactions in traffic dynamics. Because of its stochastic nature, breakdown is commonly predicted only in a probabilistic manner. This paper focuses on using stationary aggregated traffic data to capture traffic dynamics, developing and testing machine learning (ML) approaches for traffic breakdown prediction and comparing them with the traditionally used probabilistic approaches. The contribution of this study is three-fold: it explores the usefulness of temporally and spatially lagged detector data in predicting traffic flow breakdown occurrence, it develops and tests ML approaches for traffic breakdown prediction using this data, and it compares the predictive power and performance of these approaches with the traditionally used probabilistic methods. Feature selection results indicate that breakdown prediction benefits greatly from the inclusion of temporally and spatially lagged variables. Comparing the performance of the ML methods with the probabilistic approaches, ML methods achieve better prediction performance in relation to the class-balanced accuracy, true positive rate (recall), true negative rate (specificity), and positive predictive value (precision). Depending on the application of the prediction approach, the method selection criteria may differ on a case-by-case basis. Overall, the best performance was achieved by the neural network and support vector machine approaches with class balancing, and with the random forest approach without class balancing. Recommendations on the choice of prediction approaches based on the specific application objectives are also given.
ASJC Scopus subject areas
- Civil and Structural Engineering
- Mechanical Engineering