基于神经常微分存储网络的交通流量预测模型  

Traffic Flow Prediction Model Based on Neural Ordinary Differential Memory Network

作　　者：薛骐 乔少杰 彭钰寒 于泳 谢添丞 XUE Qi;QIAO Shaojie;PENG Yuhan;YU Yong;XIE Tiancheng(School of Software Engineering,Chengdu University of Information Technology,Chengdu 610225,China)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225

出　　处：《无线电通信技术》2024年第6期1215-1222,共8页Radio Communications Technology

基　　金：国家自然科学基金(62272066);四川省科技计划(25NSFSC0293,25GJHZ0297,2023YFG0027);教育部人文社会科学研究规划基金(22YJAZH088);网络空间安全教育部重点实验室及河南省网络空间态势感知重点实验室开放基金课题(KLCS20240106);成都市技术创新研发项目(重点项目)(2024-YF08-00029-GX);成都市区域科技创新合作项目(2023-YF11-00020-HZ);成都市技术创新研发项目(2024-YF05-01217-SN);大学生创新训练计划项目(202410621195,202410621183)。

摘　　要：交通流量预测是多元时空预测中的典型任务,也是智能交通系统的重要组成部分。然而,现有模型很少关注交通路网中不同道路间的共有模式。主流模型大多基于图神经网络(Graph Neural Network,GNN)实现,而GNN随着层数的增加,会出现过度平滑现象,即邻接图中表征趋于相近。为解决上述问题,提出一个神经常微分存储网络(Neural Ordinary Differential Memory Network,NODEMN),利用模式记忆单元保留时空数据中的显著特征,进行模式匹配,利用神经常微分方程改善了深度训练中出现的过度平滑问题。NODEMN在真实数据集上进行大量实验,结果表明,NODEMN模型相较于基准模型在预测性能上具有显著优势,在3个数据集上的平均百分比误差(Mean Absolute Percentage Error,MAPE)平均降低4.09%、平均绝对误差(Mean Absolute Error,MAE)平均降低3.38%、均方差误差(Root Mean Square Error,RMSE)平均降低2.49%。Traffic flow prediction is a typical task in multivariate spatio-temporal prediction and an important part of intelligent transport systems.However,existing models seldom pay attention to shared patterns among different roads in the traffic network.Most of mainstream models are implemented based on Graph Neural Network(GNN),which are subject to transition smoothing phenomenon,i.e.,the representations in the neighborhood graph tend to be similar,as the number of layers increases.In order to solve above problems,a Neural Ordinary Differential Memory Network(NODEMN)is proposed,which utilizes pattern memory units to retain salient features in spatio-temporal data for pattern matching.In addition,the over-smoothing problem that occurs in deep training is improved by using neural differential equations.NODEMN conducts a large number of experiments on real datasets.Experimental results show that the NODEMN model has a significant advantage in prediction performance compared to the baseline methods,with an average reduction of 4.09%in the Mean Absolute Percentage Error(MAPE),Mean Absolute Error(MAE)by 3.38%,and Root Mean Square Error(RMSE)by 2.49%on three datasets.

关 键 词：多元时序预测 时空数据库 图神经网络 机器学习 交通预测 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]