使用动态时空神经网络的市区交通流量预测  

Predicting Citywide Traffic Flow Using Dynamic Spatial-temporal Neural Networks

作　　者：任建华[1] 朱尧 孟祥福[1] 张霄雁[1] REN Jian-hua;ZHU Yao;MENG Xiang-fu;ZHANG Xiao-yan(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,辽宁葫芦岛125105

出　　处：《小型微型计算机系统》2023年第3期529-535,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金面上项目(61772249)资助;辽宁省教育厅一般项目(LJ2019QL017,LJKZ0355)资助。

摘　　要：准确的市区交通流量预测对交通管理、城市规划和公共安全等领域具有重要意义.现有城区交通流量预测方法主要采用CNN等深度学习模型,但存在以下问题:一方面由于捕获全局空间依赖需要堆砌很多层增加网络的接受域,导致学习全局空间依赖关系的效率低下,另一方面忽略了城市区域交通流量的动态性.针对上述问题,本文提出了一种基于注意力的动态时空神经网络市区交通流量预测模型(Spatio-Temporal 3D Convolution Global Depth Residual Network, ST-3DGN).首先,该模型使用多层三维卷积捕捉城市区域交通流动性;然后,采用改进的残差结构结合空间注意力机制对远距离区域间流的空间依赖性进行建模;最后,使用了一种早期融合机制稳定了训练过程,从而进一步提高了模型ST-3DGN的性能.在两个真实公开的数据集上进行了大量实验,实验结果表明本文提出的ST-3DGN模型在预测准确性方面明显优于现有的主流交通预测模型.Accurate urban traffic flow prediction is of great significance to the fields of traffic management, urban planning and public safety.The existing urban traffic flow prediction methods mainly adopt deep learning models such as Convolutional Neural Networks(CNNs),which mainly faced by the following two critical drawbacks: on the one hand, due to the capturing of the global spatial dependencies, they need to stack many layers to increase the receptive field of the network, leading to the inefficiency of learning the global spatial dependencies.On the other hand, the dynamics of the urban regional traffic flow is ignored.To deal with the above problems, this paper proposes an urban flow prediction model(Spatio-Temporal 3D Convolution Global Depth Residual Network, ST-3DGN)that incorporates mobility and spatiotemporal characteristics.First, the model uses multi-layer three-dimensional convolution to capture urban regional traffic mobility.And then, we use an improved residual structure combined with a spatial attention mechanism to model the spatial dependence of the flow between long-distance regions.Lastly, an early fusion mechanism is used to stabilize the training process and further improve the performance of the model ST-3DGN.Extensive experiments have been conducted on two real public available datasets and demonstrated that the proposed ST-3DGN model significantly outperforms the existing mainstream traffic flow prediction models in terms of prediction accuracy.

关 键 词：交通流量预测 时空特性 残差结构 注意力机制 融合机制 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]