引用本文: | 欧阳毅,汤文燕,邵泳博,黎晏伶.时空特征变分学习的交通流预测模型[J].控制理论与应用,2025,42(1):158~166.[点击复制] |
OUYANG Yi,TANG Wen-yan,SHAO Yong-bo,LI Yan-ling.Spatial-temporal feature variational inference model for traffic flow forecasting[J].Control Theory and Technology,2025,42(1):158~166.[点击复制] |
|
时空特征变分学习的交通流预测模型 |
Spatial-temporal feature variational inference model for traffic flow forecasting |
摘要点击 3866 全文点击 20 投稿时间:2022-12-03 修订日期:2024-09-02 |
查看全文 查看/发表评论 下载PDF阅读器 |
DOI编号 10.7641/CTA.2023.21057 |
2025,42(1):158-166 |
中文关键词 交通流预测 时空融合 变分自编码器 图卷积 |
英文关键词 traffic flow prediction spatio-temporal fusion variational autoencoder graph convolution |
基金项目 浙江工商大学“数字+”学科建设管理项目(SZJ2022C004), 浙江省智能交通工程技术研究中心项目(2015ERCITZJ–KF1)资助. |
|
中文摘要 |
交通流量时空预测是智能交通系统的关键任务. 针对城市交通流序列的非线性和多模态特性, 本文提出了一种基于时空特征融合的变分学习模型(ST-FVAE), 采用局部时空特征融合和全局特征融合两个阶段对具有图空间拓扑特性的交通流数据进行预测. 局部特征融合模块由时间卷积残差单元和图卷积神经网络(GCN)构成, 提取交通节点的局部时间特征信息, 并利用GCN将空间拓扑信息嵌入局部时间特征信息中. 通过基于局部时空图特征融合的变分自编码器交通流预测模型, 学习全局时空相关性特征. 在全局时空特征融合变分自编码器的学习过程中, 为使Q分布能够逼近实际数据P分布, 通过最大化似然函数的变分推断证据下界(ELBO)使得两个分布之间的KL散度最小化, 提出了计算分布期望的KL差异构造训练损失函数的方法, 进一步提高预测准确率. 通过对交通流数据集和交通速度数据集的预测实验结果表明: 本文提出的方法在交通流量和速度的预测方面都具有较好的预测特性, 对于30 min和60 min的预测鲁棒性更好. |
英文摘要 |
Traffic flow spatial-temporal data forecasting is a crucial task for intelligent transportation systems. This paper proposes a variational learning model based on the spatio-temporal feature fusion (ST-FVAE). The model aims to address the nonlinear and multi-modal features of urban traffic flow sequences by utilizing local spatiotemporal feature fusion and global feature fusion. It also takes into account the graph spatial topological characteristics to predict traffic flow data. The local feature fusion module is composed of a temporal convolutional residual unit and a graph convolutional neural network model (GCN). It extracts the local temporal feature information of traffic nodes and uses GCN to embed the spatial topological information into the local temporal feature information. We learn global spatial-temporal correlation features via this variational auto-encoder traffic flow prediction model of local spatial-temporal graph feature fusion. During the learning process of global spatio-temporal graph feature fusion variational auto-encoder, to make the variational Q distribution approximate the actual data P distribution, we use the variational inference ELBO (evidence lower bound) that maximizes the likelihood function to minimize the KL scatter between the two distributions. Meanwhile, we construct the training loss function using the KL function property. We perform prediction evaluation on three different large-scale traffic datasets. Experiments show that the model proposed in this paper has better prediction performance in both traffic flow and speed. Moreover, our method is more robust for 30 and 60-minute forecasting. |
|
|
|
|
|