交通拥堵是阻碍城市建设和可持续发展的主要因素之一。不仅如此，交通拥堵还会给人们的生活带来诸多不便，如长时间的通勤时间、低效的运输和资源浪费。因此，准确预测交通流对缓解拥堵现状、优化交通资源配置、制定调度策略、提升出行体验具有重要的意义。本研究对交通流预测算法进行了系统地总结，并提出两种基于时空图神经网络的交通流预测算法，分别是：（1）基于注意力机制的多图卷积递归网络（Attention-based Multiple Graph Convolutional Recurrent Network, AMGCRN），该算法能够在交通流空间特征提取过程中有效地捕捉全局的道路网络结构的相似性和交通流的局部特征。（2）时空交互图神经网络（Spatial-Temporal Interactive Graph Neural Network, STIGNN），该算法通过使用交互式学习策略来提取隐式的时空相关性。本论文的主要工作如下：（1）由于车辆流动和道路网络结构的拓扑约束，交通流具有空间相关性。本研究从图数据结构的角度出发，分别基于图卷积网络（Graph Convolutional Network, GCN）、切比雪夫网络（Chebyshev Network, ChebNet）和图注意力网络（Graph Attention Network, GAT）对交通流空间相关性建模。（2）由于车流所呈现的周期规律性，交通流具有时间相关性。本研究使用了递归神经网络的两种变体模型长短时记忆网络（Long Short-Term Memory, LSTM）、门控循环单元（Gate Recurrent Unit, GRU）对交通流进行了时间特征的提取，并且以这两种算法为基础搭建了完整的时间预测模型。（3）基于交通流的时空相关性，本研究提出基于注意力机制的多图卷积递归网络 AMGCRN。该网络能够同时提取基于路网拓扑的全局空间特征和基于流量变化的局部空间特征，然后将空间特征嵌入时序结构实现时空特征提取。AMGCRN在两个公开数据集上与多种交通流预测算法比较可实现约 1% 的效果提升。消融实验进一步证明 AMGCRN 模型结构设计的合理性。（4）基于交通流的时空交互性，本研究提出时空交互图神经网络 STIGNN。该网络引入了交互式学习策略，将交通流按照时间维度分割为两个子序列，然后通过交互策略增加有效时间特征的提取。STIGNN 在两个公开数据集上与多种最新的交通流预测算法进行比较，对比最新算法提升了约 0.5% 的预测精度。消融实验证明了 STIGNN 模型结构设计的合理性。
Traffic congestion is one of the major obstacles to urban development and sustainable growth. Moreover, traffic congestion also brings inconvenience to people’s lives, such as long commuting time, inefficient transportation, and resource waste. Therefore,accurate prediction of traffic flow is of significant importance for alleviating congestion, optimizing traffic resource allocation, formulating scheduling strategies, and enhancing travel experience. In this study, traffic flow prediction methods are comprehensively andsystematically summarized, and two traffic flow prediction algorithms based on spatio temporal graph neural networks are proposed: (1) Attention-based multiple graph convolutional recurrent network (AMGCRN), which can effectively capture the similarity of global road network structure and local features of traffic flow in the process of spatial feature extraction; (2) Spatial-temporal interactive graph neural network (STIGNN), which uses interactive learning strategies to extract implicit spatio-temporal correlations. The main contents of this thesis are as follows:(1) Due to the topological constraints of vehicle flow and road network structure, traffic flow has spatial correlation. In this study, from the perspective of graph data structure, the spatial correlation of traffic flow is modeled using graph convolutional networks(GCN), Chebyshev network (ChebNet), and graph attention networks (GAT), based on the topological constraints of road network.(2) Due to the periodic regularity presented by vehicle flow, traffic flow has temporal correlation. In this study, two variants of recurrent neural networks e.g., long short-term memory (LSTM) and gate recurrent unit (GRU), are used to extract time features of trafficflow. A complete time prediction model is built based on these two neural networks.(3) Based on the spatio-temporal correlation of traffic flow, the attention-based multiple graph convolutional recurrent network (AMGCRN) is proposed to extract both global spatial features based on road network topology and local spatial features based on trafficflow changes, and then embed the spatial features into temporal structures to achieve spatial temporal feature extraction. In this study, AMGCRN is compared with various traffic flow prediction algorithms on two public datasets for predictive performance whichachieves about 1% improvement. The ablation experiments validate the model structureof AMGCRN.(4) Based on the spatio-temporal interaction of traffic flow, the spatial-temporal interactive graph neural network (STIGNN) is proposed to enhance the extraction of effective time features through interaction strategies, which introduces interactive learningstrategies to divide traffic flow into two sub-sequences according to the time dimension. In this study, STIGNN is compared with various traffic flow prediction algorithms on two public datasets to evaluate its predictive performance which further improves the prediction performance by about 0.5%, and the scientific design of STIGNN model structure is further validated through ablation experiments.
