Abstract: Considering the facts that the spatio-temporal dependence of network traffic flow is highly complex and that traffic flow data has noises in practices, this paper proposes a novel spatio- temporal fusion model based on graph neural network for effective traffic flow forecasting. To alleviate negative impacts of data missing, data exception and data noise, a feature fusion block is designed to reconstruct input features and smoothing them within a sliding time window, and then the obtained features are fed into the main body of the proposed model. The main body adopts a design of bi-directional networks to learn respectively the forward and reverse spatio-temporal representation of traffic flow. Both networks share the same structure but with different adjacent matrices. In particular, the causal convolution is used as the temporal feature extractor, and a block of adaptive gated graph convolutional neural network is specially designed for spatial feature extracting, to realize adaptively information aggregation and propagation. Then, a lightweight longitudinal information aggregation layer is constructed to realize information fusion within different local receptive fields. At last, information contributions of the forward and reverse networks are weighed and aggregated with an attention-output module, to establish the expected Bi-directional Adaptive Gating Graph Convolutional Networks (Bi- AGGCN) for traffic flow forecasting. To validate the effectiveness of the proposed model, a series of experiments are carried out on four real traffic flow benchmark datasets, i.e., PEMS03, PEMS04, PEMS07 and PEMS08. Experimental results show that the proposed model Bi-AGGCN can outperform all baseline models over four datasets with three metrics. At the same time, compared with the state-of-the-art baselines, i.e., Spatial-Temporal Synchronous Graph Convolutional Networks (STSGCN) and Spatial-Temporal Fusion Graph Neural Networks (STFGNN), Bi-AGGCN is dramatically lighter in parameter scale and faster in training time, and achieves higher prediction accuracy at a significant lower cost.

Key words: intelligent transportation, adaptive gating unit, graph convolution, bi-directional graph networks, feature fusion, longitudinal information aggregation layer

摘要： 针对路网交通流时空依赖上的高度复杂性以及数据污染的现实性，基于图神经网络构建一种新型时空融合交通流预测模型。考虑交通流数据中的缺失、异常与噪声，模型首先对数据进行特征重构与融合，在保持时序特性的前提下，以滑动时间窗口平滑交通流特征信息，做好数据准备。考虑交通流的实际有向性，主体模型采用正、反双路网络设计以分向学习交通流时空特征的有效表示。双路网络结构相同，以轻量有效的因果卷积作为模型的时序特征提取器，以多层自适应门控图卷积神经网络作为模型组件提取空间特征，实现信息的自适应聚合与传播，再通过纵向信息聚合层轻量化地实现不同局部视野下的信息融合，基于注意力有效权衡两路网络的信息贡献并将其聚合，建立双向自适应门控图卷积网络交通流预测模型。在真实交通流基准数据集PEMS03、PEMS04、PEMS07和PEMS08上进行模型的有效性验证，结果表明，所建模型在4个数据集上3个预测精度指标均优于基线模型。同时，相较于最先进的基线模型时空同步图卷积网络与时空融合图神经网络，所建模型能以数倍甚至数十倍比例的参数轻量化与低训练时间代价获得更高的预测精度。

关键词: 智能交通, 自适应门控, 图卷积, 双向图网络, 特征融合, 纵向层间聚合