2018 Volume 5 Issue 2 Pages 349-361
Several real-time crash prediction (RTCP) models have been proposed using Bayesian networks (BNs), which are probabilistic graphical modeling methods offering a great degree of robustness. These models offer real-time applicability, high prediction success, a capacity to handle missing data, and the possibility of a flexible variable space. However, to develop an advanced RTCP model using BN, it is imperative to identify the most influential traffic variables and their combinations. This study proposes BN-based RTCP models with 24 combinations of 12 traffic variables. After modeling, their performances were validated and compared to identify the preferable combinations of input variables. The models constructed with differences between the upstream and downstream congestion index, flow, speed, and the upstream congestion index as variables proved to be the most effective combination of input variables.
