Machine learning-based accidents analysis and risk early warning of hazardous materials transportation

Abstract

In this study, we conduct a comprehensive statistical analysis of the increasing frequency of hazardous materials accidents in the U.S. highway transportation sector. Based on these findings, we propose an enhanced model designed to provide robust data support for risk warning initiatives. After analyzing over 600,000 accidents from 1971 to 2023, we observe that the annual number of accidents has exceeded 20,000 since 2021. This trend underscores the urgent need to enhance the accuracy of accident risk warnings to mitigate economic losses. The study further reveals that most accidents occur between 6:00 and 11:00, with 91.7% of these incidents resulting in spillage. This finding underscores the critical need for a robust emergency response plan specifically tailored to address spillage events. To address the issue of performance degradation of models in large-scale datasets, the “SF-T0.25” model using a stacking algorithm was developed, which was validated using more than 70,000 spillage accident records from 2021 to 2023. The results show that the prediction accuracy of the model reaches 0.9628, which is better than the parameter-adjusted ET model (0.94981). The SF-T0.25 model also performs well in indicators such as the Jaccard similarity coefficient and the cross-entropy. The mean value of Jaccard similarity coefficient in predicting the type of accident weather conditions is more than 0.97 and the mean value of Cross-Entropy Loss in predicting the range of instantaneous speed of vehicles during accidents is less than 0.05, which proves that the model can provide reliable data support for early risk warning of hazardous materials transportation.