侧向碰撞风险对交织区汇合行为的影响

为研究高速公路交织区匝道车辆的汇合行为，基于梯度提升决策树(Gradient Boosting Decision Tree, GBDT)构建了交织区汇合加速度模型，利用美国 Next Generation Simulation (NGSIM)车辆轨迹数据提取汇合车辆与周围车辆之间的横纵向距离差、速度差及加速度等交通 参数作为候选变量，将1 s后的汇合加速度作为预测变量，对模型进行五重交叉训练和测试，获取 模型最佳参数组合，引入横向距离碰撞时间 TC 分析汇合过程中侧向碰撞风险对汇合加速度的影 响。研究发现：与基于视角的刺激-反应模型(VASR)相比，本文模型的预测精度更高；引入变量 TC 在均方误差(MSE)、平均绝对偏差(MAD)和R2这3个指标上均优于原模型；在各影响变量之中，汇 合车辆与目标车道领车的速度差 ΔVPL 和横向距离碰撞时间 TC 对汇合加速度的影响最大，相对影 响程度分别达到20.2%和12.1%。研究发现，GBDT模型能够准确预测车辆汇合加速度，深入挖掘 变量和汇合加速度之间的非线性关系，引入变量 TC 能够有效提高模型精度。

Influence of Lateral Collision Risk on Merging Behavior of Weaving Area

In order to study the merging behavior of vehicles in the interweaving area of ramps of the expressway, an acceleration model of vehicle merging was constructed based on the gradient boosting decision tree (Gradient Boosting Decision Tree, GBDT). Traffic parameters such as distance difference, speed difference and acceleration, which were extracted from the NGSIM vehicle trajectory data of the United States are used as candidate variables. The combined acceleration after 1s is used as the predictor variable. Five-fold cross-training and testing are performed on the model to obtain the best parameter combination of the model. The lateral distance collision time TC is introduced to analyze the merging process and the impact of the mid-to- side collision risk on the confluence acceleration. The study indicates that: the model has higher prediction accuracy compared to the perspective- based stimulus- response model (VASR); the introduction of the variable TC performs better in the three indicators of mean square error (MSE), mean absolute deviation (MAD) and R2 than the original model; among the various influencing variables, the speed difference ΔVPL between the converging vehicle and the leading vehicle in the target lane and the lateral distance collision time TC have the greatest impact on the confluence acceleration, with the relative impact reaching 20.2% and 12.1%, respectively. The study found that the GBDT model can accurately predict the confluence acceleration of vehicles, and deeply explore the nonlinear relationship between the variables and the confluence acceleration. The introduction of the variable TC can effectively improve the accuracy of the model.