Breakdown Prediction Utilizing Data Mining Algorithms in Combination With Performance Measures Derived From Connected Vehicle Data

Abstract

Abstract Connected vehicles (CV) will provide an important source of data to support real-time management of traffic operations and off-line analysis of traffic operations. CV data allows the derivation of metrics not currently computed for use in freeway management such as the standard deviation of speed, acceleration/deceleration, and jerk. This study explores the utilization of CV data to derive such metrics for use in traffic management. The study then investigates the use of the derived metrics in combination with data analytic techniques to assess and predict the onset of congestion on freeways in real-time operations. The analysis revealed using cluster analysis and classification decision tree that the traffic states of the freeway facility used as a case study could be classified into groups representing six different traffic conditions based on speed, standard deviation of speed between vehicles, standard deviation between points, as well as the deceleration values. The study also compared the performance of three data mining/machine learning techniques for the prediction of congestion using a decision tree, a fuzzy rules-based system that utilizes the decision tree results, and a neuro-fuzzy inference system model utilizing the backpropagation optimization method for optimization. The comparison of the performance of the three prediction models demonstrated that the neuro-fuzzy inference system model achieved the best performance in terms of various performance measures that are commonly used to assess machine learning performance.