基于集成学习的交通事故严重程度预测

Traffic accident severity prediction based on ensemble learning

下载PDF

导出

摘要为提升道路交通事故严重程度预测模型的性能,以及分析事故特征对于事故严重程度的影响,提出一种基于双层Stacking模型的交通事故严重程度预测方法。首先,采用BSMOTE2算法来平衡数据,并验证数据平衡处理是否会对模型预测产生正向影响,同时利用GBDT-RFECV算法进行k折交叉验证选择,完成特征降维。其次,构建双层Stacking模型,第一层由BiGRU和XGBoost组成,将时间序列特征用于BiGRU,静态特征用于XGBoost进行初步预测;第二层采用CatBoost模型,结合第一层的预测结果进行最终的严重程度预测。研究结果表明:模型的准确率、宏F_(1)和宏AUC均有明显提高,表明数据平衡处理对模型预测产生正向影响;相较于KNN、BiGRU、RF和XGBoost模型,所提双层Stacking模型的预测准确率分别提高了5.45%、10.23%、1.78%和2.34%,宏F_(1)值提高了5.31%、9.91%、1.35%和1.92%,宏AUC提高了11.13%、6.97%、2.13%和2.71%。该双层Stacking模型在多个评估指标上的表现均优于其他模型。 In order to improve the performance of road traffic accident severity prediction models and analyze the impact of accident features on accident severity,a method of traffic accident severity prediction based on a double-layer Stacking model is proposed.The BSMOTE2 algorithm is used to balance the data and verify whether data balancing processing will have a positive impact on model prediction.The GBDT-RFECV algorithm is used for k-fold cross validation selection to complete the feature dimensionality reduction.A two-layer Stacking model is built.The first layer is composed of BiGRU and XGBoost,using time series features for BiGRU and static features for XGBoost for the preliminary prediction.The CatBoost model is used at the second layer and combined with the prediction results of the first layer for the final severity prediction.The research results indicate that the accuracy of the model,macro F_(1),and macro AUC have all improved significantly,indicating that data balance processing has a positive impact on model prediction.In comparison with KNN,BiGRU,RF,and XGBoost models,the proposed double-layer Stacking model can improve prediction accuracy by 5.45%,10.23%,1.78%,and 2.34%,respectively,the macro F1 value can be increased by 5.31%,9.91%,1.35%,and 1.92%,respectively,and the macro AUC can be increased by 11.13%,6.97%,2.13%,and 2.71%,respectively.The double-layer Stacking model can perform better than other models on multiple evaluation metrics.

作者贾现广宋腾飞吕英英 JIA Xianguang;SONG Tengfei;LÜYingying(School of Transportation Engineering,Kunming University of Technology,Kunming 650500,China;School of Information Engineering and Automation,Kunming University of Technology,Kunming 650500,China)

机构地区昆明理工大学交通工程学院昆明理工大学信息工程与自动化学院

出处《现代电子技术》北大核心 2025年第16期61-66,共6页 Modern Electronics Technique

基金国家自然科学基金项目(71961012)。

关键词交通安全交通事故预测预测分析集成学习机器学习深度学习特征降维 traffic safety traffic accident severity predictive analysis ensemble learning machine learning deep learning feature dimensionality reduction