摘要
短时交通流量预测是辅助智能交通系统进行决策,解决拥堵问题,提高道路通行能力的关键任务。为提高模型对于小样本数据的拟合效果及速度,并充分挖掘交通流序列中存在的线性与非线性关系,将对于线性数据具备良好拟合效果,并且时间复杂度较低的自回归滑动平均(ARMA)模型与对于非线性、小样本数据具有计算准确率高、时间复杂度低等优势的支持向量回归(SVR)模型进行组合,提出一种残差优化组合预测模型。采用赤池信息准则对ARMA模型进行定阶,实现交通流量的线性拟合,并得到相应的残差序列。然后将重构后的残差序列作为SVR模型的输入,对残差序列进行预测,以补偿交通流量数据中的非线性变化。将ARMA、SVR、长短期记忆网络、人工神经网络及ARMA-SVR加权组合模型作为对照组进行模型评价。结果表明:样本的时间间隔分别为5,10,15 min时,ARMA-SVR残差优化组合模型的均方根误差(RMSE)及平均绝对误差(MAE)均小于对照组模型,RMSE降低约0.378~7.063,MAE降低约0.054~0.802;ARMA-SVR残差优化组合模型在不同的样本时间间隔下均具备较高的预测能力、较低的时间复杂度及数据计算成本,可以满足基于不同样本时间间隔的交通流量预测的需要。
Short-term traffic volume forecasting is the key task to assist ITS to make decisions,solve congestion problems and improve road capacity.In order to improve the fitting effect and speed of the model for small sample data,and to fully explore the linear and nonlinear relationships existing in the traffic volume sequence,the ARMA model with good fitting effect for linear data and low time complexity is combined with the support vector regression(SVR)model with advantages of high calculation accuracy and low time complexity for nonlinear and small sample data,and an residual optimization composite forecasting model is proposed.The order of the ARMA model is set by using Akaike information criterion to realize the linear fitting of traffic volume,and the corresponding residual sequence is obtained.Then,using the reconstructed residual sequence as the input of SVR model,and the residual sequence is predicted to compensate for the nonlinear changes in traffic volume data.The ARMA,SVR,long-term and short-term memory networks,artificial neural network,and ARMA-SVR weighted composite forecasting model are used as control group for model evaluation.The result shows that(1)when the sampling time interval is 5,10,15 min,the root mean square error(RMSE)and mean absolute error(MAE)of the ARMA-SVR residual optimization composite forecasting model are smaller than those of the control models,and the RMSE and MAE are reduced by about 0.378-7.063 and 0.054-0.802 respectively;(2)the ARMA-SVR residual optimization composite forecasting model has higher prediction ability,lower time complexity and data calculation cost in different sample time intervals,which can meet the needs of traffic volume prediction based on different sample time intervals.
作者
王博文
王景升
朱茵
王统一
张泽有
WANG Bo-wen;WANG Jing-sheng;ZHU Yin;WANG Tong-yi;ZHANG Ze-you(School of Traffic Management,People's Public Security University of China,Beijing 100038,China;School of Electrical Information,Shandong University of Science and Technology,Jinan Shandong 250000,China)
出处
《公路交通科技》
CAS
CSCD
北大核心
2021年第11期126-133,共8页
Journal of Highway and Transportation Research and Development
基金
公安部公安理论及软科学研究计划项目(2020LLYJGADX020)
中国人民公安大学拔尖创新人才培养经费支持研究生科研创新项目成果(2021yjsky014)
中国人民公安大学公共安全行为科学实验室开放课题基金资助项目(2020SYS15)。
