双机制Stacking集成模型在短时交通流量预测中的应用

Application of Stacking Integrated Model Combined with Internal and External Mechanisms in Short-term Traffic Flow Prediction

交通流量预测对于智能交通管理决策具有重要意义,为克服传统单一模型预测精度低、稳定性不足,同时为解决Stacking集成模型对基学习器输出信息利用率不高的问题,提出了一种双机制Stacking集成模型。双机制包括内机制和外机制,内机制通过在元学习器中引入注意力机制来调整网络中的特征信息,外机制通过在基学习器中融入动态权重系数来调整基学习器的输出信息。通过内外结合的双机制实现对基学习器输出信息动态变动规律的挖掘和提取并增强对基学习器输出信息的利用率,从而提升模型的预测精度和稳定性。选取I5NB高速公路上的76898条数据为实证研究对象,进行了基于随机森林、GBDT(gradient boosting decision tree)和Xgboost(extreme gradient boosting)单一模型、传统Stacking集成模型及双机制Stacking集成模型的预测对比分析。实证结果证明双机制Stacking集成模型预测精度最高,验证了该模型在短时交通流量预测中的有效性。

Traffic flow prediction is of great significance for intelligent transportation management decision-making.In order to overcome the low accuracy and stability of traditional single model prediction,a dual mechanism stacking integration model was proposed to solve the problem of low utilization rate of stacking integrated model for basic learning device.Two mechanisms included internal mechanism and external mechanism.The inner mechanism adjusted the feature information in the network by introducing attention mechanism into the meta-learner,and the outer mechanism adjusted the output information of the base learner by integrating dynamic weight coefficient into the base learner.Through the combination of internal and external mechanisms,the dynamic change law of the output information of the base learner was extracted,and the utilization rate of the output information of the base learner was enhanced,so as to improve the prediction accuracy and stability of the model.The first mock exam was based on 76898 data from I5NB expressway.The prediction and comparison analysis based on random forest,gradient boosting decision tree and Xgboost(extreme gradient boosting)single model,traditional stacking integration model and stacking integration model of dual mechanism were carried out.The empirical results show that the dual mechanism stacking integrated model has the highest prediction accuracy,which verifies the effectiveness of the model in short-term traffic flow prediction.