基于动态时空神经网络模型的地铁客流预测

Metro Passenger Flow Prediction Based on Dynamic Spatio-temporal Neural Network Model

针对城市轨道交通站点客流预测问题,本文提出一种基于注意力机制的动态时空神经网络(DSTNN)模型。模型采用多分支并行架构,能够有效提取地铁客流的复杂时空特征,在空间维度上,全局和局部注意力机制相结合,实现站点间动态时空关联和静态拓扑结构的捕捉;在时间维度上,使用双向长短时记忆和注意力机制共同学习客流数据的时变规律。在杭州地铁数据集上进行实验,结果表明:相较于经典预测模型和深度学习模型,DSTNN具有更高的预测精度和训练效率。在4种不同的预测时长下,DSTNN模型平均绝对误差的平均值较基线中扩散卷积循环神经网络模型(DCRNN)和物理虚拟结合图网络模型(PVCGN)分别降低6.63%和2.57%。此外,可视化分析证明了本模型对时空关联的动态学习能力,消融实验验证了各分支的有效性。

This paper proposes a Dynamic Spatio-Temporal Neural Network(DSTNN)model based on attention mechanism for urban rail transit station passenger flow forecast.The DSTNN adopts a multi-branch parallel architecture to effectively extract the complex spatio-temporal features of metro passenger flow.In the spatial dimension,the global and local attention mechanisms are combined to capture dynamic spatio-temporal correlation between stations and static topology.In the temporal dimension,the bi-directional long short-term memory and attention mechanisms are used to learn the time-varying patterns of passenger flow data.In the experiments on Hangzhou Metro dataset,the results show that the DSTNN has higher prediction accuracy and training efficiency compared to classical prediction models and deep learning models.The average mean absolute error(MAE)over four different prediction durations is respectively 6.63%and 2.57%lower than that of the Diffusion Convolutional Recurrent Neural Network(DCRNN)and Physical-Virtual Collaboration Graph Network(PVCGN).In addition,the visualization analysis demonstrates the dynamic learning ability of this model for spatio-temporal correlations,and the ablation experiments verified the effectiveness of each branch.