A hybrid machine learning model for predicting Real-Time secondary crash likelihood

Secondary crashes usually occur within the spatio-temporal impact ranges of primary crashes, which could cause traffic disturbance and increase traffic safety problems. However, existing studies only focused on predicting the likelihood of crashes leading to secondary crashes without considering the likelihood of the occurrence of secondary crashes. In addition, previous studies did not consider the real-time implementation of secondary crash likelihood prediction models and included too many features that were not available in real-time. A real-time secondary crash likelihood prediction model aims to predict the likelihood of secondary crashes in a short period (e.g., 5-10 min) and update every minute. The main objective of this paper is to develop a machine learning model to predict secondary crash likelihood in real-time. Two XGBoost models were developed for predicting the likelihood of crashes leading to secondary crashes and the likelihood of the occurrence of secondary crashes, respectively. A hybrid model was proposed to combine the results from the two developed models. Results indicated that the proposed hybrid model significantly improved the accuracy of secondary crash likelihood prediction. The proposed model has the potential to be applied in proactive traffic safety management system and prevent the occurrence of secondary crashes. Moreover, experimental results suggested that several features related to real-time traffic flow conditions were crucial for predicting secondary crash likelihood, such as the average traffic volume and average occupancy.