Outlier Traffic Flow Detection and Pattern Analysis Under Unplanned Disruptions: A Low-Rank Robust Decomposition Model

Outlier traffic flow detection under unplanned disruptions is vital for operational safety and management. Though numerous models have been proposed to effectively detect outlier traffic flow in previous studies, it is also essential to simultaneously estimate the traffic flow change caused by the unplanned disruption since the information supports the dispatch of traffic resources. To fill this research gap, the study aims to decompose the observed traffic flow matrix under unplanned disruptions into a regular flow matrix and an outlier flow matrix (i.e., flow change matrix). Specifically, we develop a low-rank robust decomposition model by formulating the decomposition problem as a convex optimization problem to minimize the recovering errors of the regular flow and flow change from the observed flow. We further consider the random fluctuations of regular flow and restrict the recovering errors concerning the magnitude of random fluctuation. The accelerated proximal gradient solution algorithm is applied to obtain the decomposition results. The spatiotemporal patterns of detected outlier traffic flow are analyzed at a station level. Extensive experiments are conducted to validate the model using synthetic and real-world data from a Special Administrative Region of China. The results show that the proposed model consistently outperforms the baseline model and is more robust to traffic flow noise during rush hours. The outlier flow pattern analysis shows that the outlier entry flow and exit flow exhibit four and five patterns at a station level, which is mainly influenced by the characteristics of unplanned disruptions and the response strategies, including the blocked segment, location, time, and such as shuttle bus, detrainment, temporal closure of the stations. The findings enable operators to make proactive preparations when a new disruption occurs.