Abstract: In order to reduce highway traffic congestion and accident casualties and property losses, it is necessary to improve the accuracy and applicability of accident duration prediction results. Based on the data of 1 362 highway traffic accidents, 16 factors affecting the duration of highway traffic accidents were selected as characteristic variables. After statistical analysis of continuous characteristic variables and assignment of discrete characteristic variables, a prediction model for the duration of highway traffic accidents based on SO-BiLSTM was constructed. The research results show that the mean values of average traffic flow, average vehicle speed, and speed deviation are the smallest within the roadway sections with accident durations greater than 120 min, which are 27 882 pcu/h, 90.4 km/h, and 18.0 km/h, respectively, and the mean value of the large vehicle mixing rate is the smallest within the roadway sections with accident durations of less than 30 min, which is 34.0%. The average age of the perpetrators with an accident duration of ［60, 90) min is the largest, which is 45 years old. The average driving experience of the perpetrator with an accident duration greater than 120 minutes is the highest, which is 91 months. The number of iterations and the population size of the proposed SO-BiLSTM model are set to 40 and 30 as the optimal. The MAPE value for predicting the duration of corresponding accidents is 8.9%, which is 1.7% to 7.6% lower than PSO Elman, BiLSTM CNN, GA-BP, and LSTM, and improves the accuracy of the prediction result for accident durations more than 120 minutes. The research results are helpful to formulate highway traffic accident relief control and emergency rescue measures and improve the level of highway traffic safety management.

Key words: traffic engineering; traffic safety; highway; traffic accident duration; snake optimizer; BiLSTM

摘要： 为减少高速公路交通拥堵和事故伤亡程度及财产损失，提高事故持续时间预测结果精度和适用性，基于1 362起高速公路交通事故数据，甄选16个高速公路交通事故持续时间影响因素作为特征变量。通过对连续特征变量统计分析和离散特征变量赋值后，构建基于SO-BiLSTM的高速公路交通事故持续时间预测模型。研究结果表明，事故持续时间大于120 min的路段内平均交通流量、平均车速和车速离差等均值最小，分别为27 882 pcu/h、90.4 km/h和18.0 km/h;事故持续时间小于30 min的路段内大型车混入率均值最小，为34.0%;事故持续时间为［60，90) min的肇事者年龄均值最大，为45岁;事故持续时间大于120 min的肇事者驾龄均值最大，为91月。SO-BiLSTM模型的迭代次数和种群规模分别设置为40和30为最优，对应的事故持续时间预测结果MAPE值为8.9%，相较于PSO-Elman、BiLSTM-CNN、GA-BP和LSTM等降低1.7%~7.6%，且提高了事故持续时间大于120 min的预测结果精度。研究结果有助于制定高速公路交通事故疏解管控和应急救援措施，提升高速公路交通安全管理水平。

关键词: 交通工程；交通安全；高速公路；交通事故持续时间；蛇群优化算法；双向长短时神经网络