考虑多维动态特征交互的高速公路实时事故风险建模

This paper investigates the impact of weather, road features, and the dynamic mutual interactions among traffic flow, weather, road, and time on the accuracy of real-time crash risk prediction. The study developed four datasets based on the crash data, traffic sensor data, weather data, and road data collected from the Beijing section of the Beijing-Harbin Freeway. The datasets include (1) the simple traffic flow data; (2) the combined traffic flow, weather, and time data; (3) the combined traffic flow, road, and time data; (4) combined traffic flow, weather, road, and time data. By considering the interactions of multi-dimensional dynamic features, this study proposes a real-time crash risk prediction model based on the Deep & Cross Network (DCN). The results demonstrate that the DCN model achieves higher accuracy than other methods in real-time crash risk prediction. The Area Under Curve (AUC) of the model is 0.8562 and the proposed model is able to correctly classify 84.26% of non-crash data and 77.55% of crash data with the probability threshold of 0.2. The DCN model used in this study can effectively predict the occurrence of freeway crashes and collisions in time, under the condition of multi-dimensional dynamic feature interactions. The proposed method has great potential to support the freeway safety management departments of China in both theoretical and technical aspects.