自适应多视图融合图神经网络地铁客流预测模型

Metro Passenger Flow Prediction Model Using Adaptive Multi-view Fusion Graph Neural Network

针对传统方法对地铁车站的多视角空间交互建模不足的问题,本文提出自适应多视图融合图神经网络模型(Adaptive Multi-view Fusion Graph Neural Network Model, AMFGNN)进行地铁车站短时客流预测。在空间维度,模型包括了物理拓扑图、线路可达性图、空间距离图等多个局部视图,并使用图注意力网络(Graph Attention Networks, GAT)学习每个视图内车站间的动态空间交互;以单视图车站为中心节点,结合其他视图中该车站作为邻居节点构建融合视图,并使用GAT学习多视图间动态交互;在时间维度,使用门控循环单元神经网络学习车站客流的时变特征。以重庆市地铁网络为例,全网出站客流的预测实验结果表明:相较于基线中的物理虚拟结合图网络模型(PVCGN),AMFGNN的平均绝对误差和均方根误差分别降低3.06%和2.49%。多视图内节点间注意力分数可视化结果表明,基于GAT的多视图建模思路能够自适应地融合不同视图中提取到的车站空间信息。此外,AMFGNN模型性能影响因素分析结果表明,以物理拓扑、线路可达性等结构稳定的视图作为中心节点构建融合视图能够获得更准确、稳定的预测模型。

To solve the problem of insufficient modeling of multi-view spatial interaction in metro stations by traditional methods,this study proposes an Adaptive Multi-view fusion Graph Neural Network Model(AMFGNN)to conduct spatial interaction modeling in metro stations short-term passenger flow prediction.In the spatial dimension,the model includes multiple partial views such as physical topology graph,line accessibility graph,spatial distance graph,etc.,and uses the graph attention networks(GAT)to learn the dynamic spatial interaction within a single view.Taking the single-view station as the central node,combined with the station in other views as neighbor nodes,this paper constructs a fused view is and uses the GAT is to learn the dynamic interaction between multiple views.In the time dimension,the gated recurrent unit neural network is used to learn the time-varying characteristics of station passenger flow.The experiments were conducted in the Chongqing metro network,and the prediction results of the outbound passenger flow of the entire network show that compared with the physical virtual combined graph network model(PVCGN)in the baseline,the AMFGNN can reduce the average absolute error and root mean square error of the network's outbound passenger flow respectively by 3.06%and 2.49%.The visualization results of attention scores between nodes in multi-views graph show that the multi-view modeling based on the GAT can adaptively and effectively integrate station spatial information extracted from different views graph.In addition,the analysis of the impact factors of AMFGNN model performance show that using structurally stable views graph such as physical topology and line accessibility as central nodes to build a fusion view graph can obtain a more accurate and stable prediction model.