摘要
【目的】为提高高速路合流区安全水平并揭示交通冲突发生机理,研究合流区交通冲突预测及成因问题。【方法】通过整合Exid高精度轨迹数据与Lanelet2高清地图,构建多维时空特征数据库。运用互信息(MI)、XGBoost和GPT算法生成多角度独立特征集,并构建残差卷积神经网络(ResCNN)预测交通冲突。通过生成混淆矩阵对预测结果进行可视化呈现;利用Accuracy、Recall等多项指标,对比ResCNN与CNN、AttCNN、ConvXGB、Transformer、GraphSAGE在不同特征集下的综合性能;并进一步基于Friedman-Nemenyi开展模型差异显著性分析;利用AUC验证冲突事件检测能力并确定最优特征集;引入SHAP算法分析交通冲突单特征贡献度与双特征交互效应。【结果】预测结果可视化下,混淆矩阵表明ResCNN能准确识别绝大部分冲突事件,且误判率较低。综合性能评价下,ResCNN在4种特征集下均表现优异,各项指标均超过93.5%,特别是在GPT&XGB_selector特征集下,其准确率高达99.27%,召回率达到99.03%,性能接近理论极限。显著性检验下,显著性概率P值远小于置信水平,且ResCNN与其他模型的平均位次差基本都大于临界值。检测能力验证下,ResCNN对应曲线在各特征集下均保持最陡峭上升趋势。模型可解释性分析下,单特征贡献度明确了车头时距等8个关键特征对冲突的不同影响效果;双特征交互归因呈现出车头时距与速度差等多组特征的复杂关联。【结论】ResCNN相比其余模型存在显著性优势,能精确区分冲突与非冲突事件,且适配于不同特征集,有效解决了高速路合流区冲突预测与机理解析问题,为智能交通系统的冲突预测提供了新的解决方案。
[Objectives]To enhance the safety of highway merging zones and uncover the mechanisms underlying traffic conflicts,this study investigates traffic conflict prediction and contributing factors in merging areas.[Methods]A multi-dimensional spatiotemporal feature database was constructed by integrating high-precision Exid trajectory data with Lanelet2 HD maps.Mutual Information(MI),XGBoost,and GPT algorithms were employed to generate multi-perspective independent feature sets.A Residual Convolutional Neural Network(ResCNN)was then developed for traffic conflict prediction,with predictive outcomes visualized using a confusion matrix.Performance metrics including Accuracy and Recall were used to compare ResCNN with CNN,AttCNN,ConvXGB,Transformer,and GraphSAGE models across different feature sets.The Friedman-Nemenyi test was conducted to assess the statistical significance of model performance differences.The Area Under the Curve(AUC)was used to evaluate conflict detection capability and determine the optimal feature set.The SHAP(SHapley Additive exPlanations)algorithm was applied to analyze both single-feature contributions and dual-feature interaction effects on traffic conflicts.[Results]Visualization of the prediction results via confusion matrices demonstrated that ResCNN accurately identified the majority of conflict events with low misclassification rates.In a comprehensive performance evaluation,ResCNN outperformed all comparative models across four feature sets,with all metrics exceeding 93.5%.Under the GPT&XGB_selector feature set,it achieved near-theoretical-limit performance,with an accuracy of 99.27%and a recall of 99.03%.Significance testing confirmed ResCNN's statistically superior performance(p-value<<confidence level),with its average rank difference exceeding critical values in most comparisons.In detection capability validation,ResCNN's AUC curve showed the steepest ascent across all feature sets.Interpretability analysis revealed:(1)Single-feature contributions highlighted eight key factors(e.g.,time headway)with distinct influence patterns;(2)Pairwise-feature interactions uncovered complex relationships between variables such as time headway and speed difference.[Conclusions]ResCNN demonstrates statistically significant advantages over comparative models,accurately distinguishing between conflict and non-conflict events while maintaining adaptability to different feature sets.The model effectively addresses both prediction and mechanistic analysis of traffic conflicts in highway merging zones,offering a novel solution for conflict prediction in intelligent transportation systems.
作者
曹弋
齐浩轩
赵斌
CAO Yi;QI Haoxuan;ZHAO Bin(School of Transportation Engineering,Dalian Jiaotong University,Dalian 116028,China)
出处
《地球信息科学学报》
北大核心
2025年第8期1841-1857,共17页
Journal of Geo-information Science
基金
辽宁省属本科高校基本科研业务费专项资金资助项目(LJ212410150048)
2023年度辽宁省研究生教育教学改革研究项目(LNYJG2023138)。
