Recurrent and non-recurrent bottleneck analysis based on traffic state rank distribution

ABSTRACTRoad network bottleneck would shift around according to the demand or supply, and therefore behaves differently from static bottlenecks such as on–off ramp and lane drop. The identification of dynamic bottlenecks is crucial for management and control. Traditional methods use direct measurements. Such methods may be biased and can only analyze the recurrent or non-recurrent bottleneck, not both. This research uses rank measurement based on speed to identify bottlenecks. The underlying idea is that bottleneck links or areas have relatively lower rank, and neighboring links covered by the same congestion event should behave similarly with respect to the rank distribution. The identification procedure is implemented by three steps: firstly, the rank distribution is fit using Gaussian Mixture Model (GMM), and then the Kullback–Leibler (KL) divergence is employed to measure the similarity between adjacent links; at last, the bottleneck areas are clustered. Using rank information would enhance the bottleneck identification performance.