城市快速路匝道车辆汇入影响因素识别与行为预测

分析了美国US101快速路瓶颈路段(简称“US101”) 和上海市延安高架上虹许路匝道瓶颈路段(简称“SHHX”) 的车辆轨迹数据, 研究了城市快速路入口匝道车辆的汇入行为; 考虑了汇入行为的13个瞬时影响因素和25个历史经历影响因素, 采用随机森林算法对38个影响因素进行重要度排序, 并识别关键影响因素; 分别对2个瓶颈路段构建了贝叶斯网络汇入行为预测模型, 并评价了模型预测精度。分析结果表明: 瓶颈路段US101和SHHX共有14个关键影响因素, 包括6个历史经历影响因素, 其中瓶颈路段US101和SHHX历史经历影响因素分别占关键影响因素总数的45.45%、36.36%, 说明汇入行为的历史经历影响因素对最终汇入决策有显著的影响; 考虑历史经历影响因素的贝叶斯网络模型预测精度较高, 瓶颈路段US101和SHHX汇入行为的总体预测精度分别提高了2.53%、8.85%, 其中未考虑历史经历影响因素时, 瓶颈路段US101和SHHX汇入行为的总体预测精度分别为87.94%和73.17%, 考虑历史经历影响因素时, 瓶颈路段US101和SHHX汇入行为的总体预测精度分别为90.47%和82.02%;此外, 预测模型无过度拟合, 测试集汇入事件与非汇入事件的预测精度之差在1.2%以内。 

Merging influence factors recognition and behaviors prediction of on-ramp vehicles of urban expressway

The vehicle trajectory data of bottleneck sections on US Highway 101 (US101) and Hongxu Road of Yan-an Expressway in Shanghai (SHHX) were analyzed, and the merging behaviors of on-ramp vehicles on urban expressway were studied.13 instant influence factors and 25 historical experienced influence factors of merging behaviors were considered. The random forests algorithm was used to sort the importances of 38 influence factors, and the key influence factors were identified. The merging behaviors prediction models based on the Bayesian network were established for two bottleneck sections, and the model prediction accuracy was evaluated. Analysis result shows that 14 key influence factors are obtained in the bottleneck sections US101 and SHHX, including 6 historical experienced influence factors. For the bottleneck sectionUS101, the historical experienced influence factors account for 45.45% of the whole key influence factors, while for SHHX, the proportion is 36.36%. Thus, the historical experienced influence factors have significant effect on the decision of final merging. The Bayesian network models that consider the historical experienced influence factors have higher prediction accuracy. The overall prediction accuracies of merging behaviors improve by 2.53% and 8.85% at US101 and SHHX, respectively. Without considering the historical experienced influence factors, the overall prediction accuracies of merging behaviors at US101 and SHHX are 87.94% and 73.17%, respectively. Under considering the historical experienced influence factors, the overall prediction accuracies of merging behaviors at US101 and SHHX are 90.47% and 82.02%, respectively. Additionally, the prediction models are not overly fitted. The difference between the prediction accuracies of merging and nonmerging events is within 1.2% using the testing datasets.