基于深度学习的短期交通流预测方法综述

Review on short-term traffic flow prediction methods based on deep learning

针对基于深度学习的短期交通流预测问题,揭示了时空相关性建模本质,分析了建模过程中涉及的多尺度时空特性、异质性、动态性、非线性等特点,明确了基于深度学习进行短期交通流预测的核心挑战,阐述了短期交通流预测涉及的外部信息整合、多步预测与单步预测以及单体预测与集成预测等相关问题;按照网格化和拓扑化2种交通流数据组织方式,分别综述了当前最新的基于深度学习的短期交通流预测研究方向。研究结果表明:针对网格化交通流数据,当前研究主要包含了基于2D图像卷积神经网络、基于2D图像卷积神经网络与循环神经网络相结合、基于3D图像卷积神经网络3种预测建模方法;针对拓扑化交通流数据,当前研究主要包含了基于1D因果图像卷积与卷积图神经网络相结合、基于循环神经网络与卷积图神经网络相结合、基于自注意力与卷积图神经网络相结合、基于卷积图神经网络的时空同步学习4种预测建模方法;总体上,基于深度学习方法进行短期交通流预测相较于采用时间序列和经典机器学习方法获得了预测准确性上的极大提升;未来,针对物理理论、知识图谱与深度学习相结合,构建多时空数据挖掘大模型以及轻量化、可解释性、模型结构自动化搜索等维度的相关探索将成为重要研究方向。

For the short-term traffic flow prediction problem based on deep learning,the essence of spatiotemporal correlation modeling was revealed,the multi-scale spatiotemporal characteristics,heterogeneity,dynamics,nonlinearity,and other characteristics involved in the modeling process were analyzed,the core challenges were clarified,and the external information integration,multi-step prediction and single-step prediction,as well as individual prediction and integrated prediction were elaborated.The latest research directions were reviewed with two organization methods of traffic flow data:grid and topology.Research results indicate that for gridded traffic flow data,current research mainly includes three prediction modeling methods:2D image convolutional neural network,2D image convolutional neural network combined with recurrent neural network,and 3D image convolutional neural network.For topological traffic flow data,current research mainly includes four prediction modeling methods:1D causal image convolution combined with convolutional graph neural network,recurrent neural network combined with convolutional graph neural network,self-attention combined with convolutional graph neural network,and spatiotemporal synchronous learning of convolutional graph neural network.Overall,the short-term traffic flow prediction based on deep learning methods significantly improves the prediction accuracy compared to time series method and classical machine learning method.In the future,the combination of physics theory,knowledge graphs,and deep learning,the construction of large-scale models for multi-temporal and spatial data mining,as well as the lightweight,interpretability,and automated model structure search,will become important research directions.1 tab,8 figs,81 refs.