Predicting crash frequency using an optimised radial basis function neural network model

Abstract

With the enormous losses to society that result from highway crashes, gaining a better understanding of the risk factors that affect traffic crash occurrence has long been a prominent focus of safety research. In this study, we develop an optimised radial basis function neural network (RBFNN) model to approximate the nonlinear relationships between crash frequency and the relevant risk factors. Our case study compares the performance of the RBFNN model with that of the traditional negative binomial (NB) and back-propagation neural network (BPNN) models for crash frequency prediction on road segments in Hong Kong. The results indicate that the RBFNN has better fitting and prediction performance than the NB and BPNN models. After the RBFNN is optimised, its approximation performance improves, although several factors are found to hardly influence the frequency of crash occurrence for the crash data that we use. Furthermore, we conduct a sensitivity analysis to determine the effects of the remaining input variables of the optimised RBFNN on the outcome. The results reveal that there are nonlinear relationships between most of the risk factors and crash frequency, and they provide a deeper insight into the risk factors’ effects than the NB model, supporting the use of the modified RBFNN models for road safety analysis.