Freeway’s Traffic Flow Breakdown Prediction Utilizing Disturbance Metrics Based on Trajectory Data

There have been limited efforts to investigate the potential of using detailed trajectory data obtained from connected vehicles and/or other sensors in deriving measures for use in real-time traffic state estimation. This study utilizes a hybrid machine learning approach that classifies the traffic states as a function of traffic disturbance and safety surrogate metrics estimated based on detailed trajectories combined with macroscopic traffic metrics. The investigated disturbance metrics are the number of oscillations, and a measure of disturbance duration based on the time exposed time to collisions. The study, first, used unsupervised clustering techniques to classify traffic states into “breakdown” and “non-breakdown” in terms of both mobility and safety. Then, the categorized traffic state was used as a binary response to the macroscopic and microscopic metrics, as features, to train supervised machine learning techniques for predicting traffic flow breakdown in the following 5-min interval in real-time operations. The study found that the utilizing disturbance and safety surrogate metrics in the real-time classification of traffic flow state increases the accuracy of prediction.