PREDICTING CRASHES BASED ON ARTIFICIAL NEURAL NETWORKS AND IDENTIFYING THE HAZARDOUS CRASH TYPE AT INTERSECTIONS

摘要

Research works to predict the traffic crash and to identify the hazard plane intersections are reviewed firstly. Then, this paper studies the use of a well-known artificial neural network (ANN), the multilayer perceptron (MLP), in predicting the number of each type of crash. The crash data include 4340 fatal/injury crashes occurred on 197 at-grade intersections in Harbin city from 2000 to 2004. Modeling result showed that the MLP is capable of predicting the crashes by type, and the testing accuracy is up to 89%. Lastly, a method to identify the hazardous crash patterns for a single intersection is presented based on the "overrepresented" concept.