基于多元时空关系的公路交通流量时序预测模型

A Model for Time Series Prediction on Highway Traffic Flow Based on Multivariate Spatio-temporal Relationship

为缓解高速公路养护施工作业对高速公路车辆通行的影响,以甘肃省酒泉高速公路为研究对象,提出了一个基于多元时空关系的公路交通流量时序预测模型。首先,使用数据结构化定义模块将一维时序数据结构化建模为二维特征数据,便于提取序列数据的空间关系。随后,将特征送入自注意力特征提取模块,使模型学习数据样本的时空关系依赖,并根据数据分布对有利于预测任务的特征分配更大的权重。最后,将两个公开交通流数据集(PEMS4与PEMS8)与甘肃省真实半封闭养护施工路段的交通流量数据集进行对比试验,计算出时序预测模型与其他模型在PEMS4与PEMS8上的数据预测误差。结果表明:时序预测模型在两个交通流数据集上有着优秀的性能表现;对比当前SOTA算法图神经网络STFGNN,时序预测模型在PEMS4数据集上的MAE减少0.53,RMSE减少0.45,在PEMS8数据集上MAE减少0.83,RMSE减少0.96;在甘肃省酒泉高速公路真实的半封闭养护施工路段的交通流量数据上,时序预测模型在真实交通流量预测任务中有着更低的预测误差,比STFGNN的MAE减少1.19,达到0.83,RMSE减少2.26,达到2.14,表明时序预测模型在交通流预测中有着更好的预测精度。研究结果可为甘肃省高速公路养护施工提供准确的交通流预测。

In order to alleviate the influence of highway maintenance work on expressway vehicle traffic,taking Jiuquan expressway in Gansu province as the study object,a time series prediction model of expressway traffic flow based on multivariate spatio-temporal relations is proposed.First,one-dimensional time series data is modeled as two-dimensional feature data by using data structure definition module,which is convenient to extract the spatial relationship of sequence data.Then,the feature is fed into the selfattention feature extraction module,which makes the model learn the spatiotemporal dependence of the data samples and assign more weight to the features that are beneficial to the prediction task according to the data distribution.Finally,two open traffic flow datasets(PEMS4 and PEMS8)are compared with the real traffic flow datasets of semi-closed maintenance road sections in Gansu province,and the data prediction errors of PEMS4 and PEMS8 between time series prediction model and other models are calculated.The result shows that(1)the time series prediction model has excellent performance on two traffic flow data sets;(2)compared with STFGNN,the MAE and RMSE decreased 0.53 and 0.45 respectively on PEMS4 and 0.83 and 0.96 respectively on PEMS8;(3)on the real traffic flow data of the semi-closed maintenance section of Jiuquan expressway in Gansu province,the time series prediction model has a lower prediction error in the real traffic flow prediction task,which is 1.19 less than MAE of STFGNN,the RMSE decreases by 2.26 and reaches 0.83 and 2.14 respectively,indicating that the time series prediction model has better prediction precision in traffic flow prediction.The study result can provide accurate traffic flow forecast for expressway maintenance construction in Gansu province.