摘要
为了进一步提升城市绕城高速收费站交通流的预测精度,根据收费站交通流的特点,该文提出了一种基于变分模态分解(VMD),长短时记忆网络(LSTM)以及支持向量回归(SVR)的城市绕城高速收费站短时交通流量预测模型,并引入遗传算法对模型中的LSTM模块和SVR模块相应的输入参数进行优化。模型中的VMD模块将原始交通流量序列分解为具有不同频率的时间序列分量和随机分量。其中第一个主要分量和随机分量分别反映了原始交通流序列的整体日变化趋势、季节变化趋势和随机干扰信息,其他分量反映了具有不同频率周期的交通流变化规律。然后,引入LSTM和SVR模型分别对不同分量进行逐一预测,并将所有分量的预测值相应叠加从而得到原始交通流量序列的最终预测结果。基于贵阳市绕城高速收费站联网收费数据,选取贵阳北主线、金华主线、上麦、贵阳东、曹关、贵阳南、石板哨、秦棋等8个收费站进行了实例验证,发现相比于KNN、BP、SVR、LSTM和ARIMA模型,在同一参数组合下,VMD-LSTM-SVR模型的MAPE均值为11.30%,且MAE和RMSE均为最低,R~2和Accuracy均为最高。这表明,提出的VMD-LSTM-SVR模型不仅具有良好的预测性能,而且还具有良好的泛化性。
In order to improve prediction accuracy of traffic flow at existing high-speed toll stations,according to the characteristics of the traffic flow of the toll station,a short-term traffic flow prediction model for urban ring expressway toll stations based on variational mode decomposition(VMD),long short-term memory(LSTM)networks,and support vector regression(SVR)is proposed,and the input parameters of LSTM module and SVR module in the model are optimized by genetic algorithm.The VMD module in the model decomposed the original traffic flow sequence into time series components with different frequencies and random component.Among them,the first major component and the random component reflect the overall daily variation trend,seasonal variation trend and random disturbance information of the original traffic flow sequence respectively,while the other components reflect the traffic flow variation law with different frequency periods.Then,LSTM and SVR models are introduced to predict the different components,and the predicted values of all components are superimposed accordingly to obtain the final prediction results of the original traffic flow sequence.Based on the toll collection data of Guiyang City Ring Expressway toll stations,eight toll stations including Guiyang North Main Line,Jinhua Main Line,Shangmai,Guiyang East,Caoguan,Guiyang South,Shibanshao,and Qinqi were selected for case analysis.Under the same set of model parameters and compared with the KNN,BP,SVR,LSTM and ARIMA models,the average MAPE of the VMD-LSTM-SVR model is 11.30%,which is at least an average reduction of 9.95%(13.86%),and the MAE and RMSE are the lowest,and the R~2 and Accuracy are the highest.This shows that the proposed VMD-LSTM-SVR model not only has good prediction performance,but also has good generalization.
作者
帅春燕
杨锰
高伦
邹辉
SHUAI Chunyan;YANG Meng;GAO Lun;ZOU Hui(Faculty of Transportation Engineering,Kunming University of Science and Technology,Kunming 650500,China;Yunnan Design&Research Institute,Tianjin Municipal Engineering Design&Research Institute Co.Ltd.,Kunming 650224,China)
出处
《测绘科学》
CSCD
北大核心
2024年第2期175-186,共12页
Science of Surveying and Mapping
基金
国家基金地区基金项目(62362044)。
关键词
智能交通
收费站短时交通流量预测
变分模态分解
长短时记忆神经网络
支持向量回归
遗传算法
intelligent transportation
short-term traffic flow prediction at toll station
variational mode decomposition
long short-term memory neural networks
support vector machine for regression
genetic algorithm
