摘要
经济间的紧密联系使城市群区域间多相关收费站交通数据间存在空间关联,准确描述该联系对提升高速公路交通流预测的精度具有积极作用.然而,由于受诸多不确定性的影响,该关联性难以捕捉和量化.针对这一缺陷,该文提出一种ATGCN-ResGRU深度学习高速公路交通流预测方法.通过结合注意力机制构建高、中、低注意度的3个GCN拓扑网络,并根据各个网络的注意度加权获得空间学习数据,将多相关收费站的联系进行量化和分级;同时,为了避免过平滑问题,将两个GRU(Gated Recurrent Unit)模块通过残差连接,进一步提升算法捕捉时间规律的能力;最后,使用特征融合层和全连接层输出预测值.利用该算法预测广东省某高速收费站的交通流量,试验结果表明:该文提出的方法能够有效提升预测精度,与经典模型多元集成CNN-LSTM、CNN-BiLSTM和DL-SVR相比,平均绝对误差(EMAE)分别减小了7.95、4.52、12.88,均方根误差(ERMSE)分别减小了12.03、6.12、19.05.
Abstract
The strong inter-economic connection makes a spatial correlation between traffic data from multiple related toll stations between urban cluster regions,and an accurate description of this connection can improve the accuracy of expressway traffic flow prediction.However,due to many uncertainties,the correlation is difficult to be captured and quantified.To solve this problem,an ATGCN-ResGRU deep learning-based expressway traffic flow prediction method was proposed.By combining attention mechanisms,three graph convolutional networks(GCN)topological networks with high,medium,and low attention levels were constructed,and spatial learning data was obtained according to the weighted attention level of each network.The connection of multiple related toll stations was quantified and graded.At the same time,to avoid the over-smoothing problem,two gated recurrent unit(GRU)modules were connected by residuals to enhance the algorithm's ability to capture time regularity.Finally,a feature fusion layer and a fully connected layer were used to output the predicted values.This algorithm was used to predict the traffic flow at a expressway toll station in Guangdong Province,and the experimental results show that the method proposed in this paper can effectively improve the prediction accuracy.Compared with the classical models of diverse ensemble CNN-LSTM,CNN-BiLSTM,and DL-SVR,the mean absolute error(EMAE)is reduced by 7.95,4.52,and 12.88,and the root means square error(ERMSE)is reduced by 12.03,6.12,and 19.05.
基金项目
国家自然科学基金资助项目(61976055)
福建省自然科学基金高校联合资助项目(2023J01946)
国家科技期刊平台
NSTL
万方数据