Abstract: Robust crash prediction is critical for deploying traffic law enforcement and emergency rescue resources in advance. So far, real-time crash prediction works are mostly at 5-minute intervals, and their results are oriented toward proactive traffic safety management of intelligent transportation systems but are too pressing for manual traffic safety management. Therefore, this study attempts to conduct the hourly traffic crash prediction to give relevant departments enough time to take measures in advance. A freeway portion in Shanghai was chosen and separated into homogenous segments, with meteorological data, traffic operation data, and crash data collected for each hour, resulting in an imbalanced dataset of crash and non-crash. To deal with the large imbalanced dataset and produce high crash prediction accuracy, an AdaBoost-CNN model, which is an integration of Adaptive Boosting and Convolutional Neural Network, was employed. The Extreme Gradient Boosting (XGBoost) and Random Forest models were also trained based on resampled datasets by Synthetic Minority Over-sampling Technique (SMOTE) and compared with the AdaBoost-CNN model. The XGBoost and Random Forest models turned out to have a poor performance on hourly crash prediction even though their training datasets were resampled by SMOTE. In addition, comparing with previous papers, the classic SMOTE method is not enough to deal with the extremely imbalanced issue. The AdaBoost-CNN model that trained through the dataset resampled by SMOTE, however, outperformed the other models in the present study and the models in similar previous research, indicating that the AdaBoost-CNN method has the potential to deal with imbalanced crash data.
Mingjie Feng, Xuesong Wang*, Bowen Cai, Ahmad Yehia, Minghui Zhong. Hourly Traffic Crash Prediction using Environmental and Electric Vehicle Big Data. Transportation Research Board 101th Annual Meeting, Washington D.C., USA, 2022.1.9-13.