时空动态图卷积架构赋能交通速度预测问题

Empowering Traffic Speed Prediction via Spatio-Temporal Dynamic Graph Convolutional Architecture

摘要: 精准的交通速度预测是智能交通系统造福人类出行的前提，但现有研究对于交通速度的预测较少，且难以捕捉与融合路网空间特征和交通速度的时序特征，导致预测的精度以及泛化能力不足。为充分捕捉交通速度的时序特征和路网的空间特征，本研究提出一种基于Transformer与图神经网络的STiGHT架构进行时空特征的捕捉，结合预定义静态图与融合时序特征的动态图进行建模，利用门控机制融合时空特征进行建模，对路网交通速度进行预测。基于在真实路网数据METR-LA上的预测任务验证了STiGHT模型预测精度显著优于基线模型，误差结果对比表明了该模型在实际应用上的可行性。STiGHT模型通过结合动态图、静态图、Transformer架构以及门控机制更好地融合捕捉了时空特征，提高了路网交通速度的预测精度以及稳定性。

Abstract: Accurate traffic speed prediction is a prerequisite for intelligent transportation systems to benefit human travel. However, existing research on traffic speed prediction is limited and struggles to capture and integrate the spatial characteristics of the road network and the temporal characteristics of traffic speed, resulting in insufficient prediction accuracy and generalization ability. To fully capture the temporal characteristics of traffic speed and the spatial characteristics of the road network, this study proposes a STiGHT architecture based on Transformer and graph neural networks to capture spatiotemporal features. It combines predefined static graphs with dynamic graphs that integrate temporal features for modeling and uses a gating mechanism to fuse spatiotemporal features for modeling, thereby predicting the traffic speed of the road network. The prediction task on the real road network data METR-LA verified that the prediction accuracy of the STiGHT model is significantly better than that of the baseline model. The comparison of error results indicates the feasibility of this model in practical applications. The STiGHT model better integrates and captures spatiotemporal features by combining dynamic graphs, static graphs, the transformer architecture, and the gating mechanism, thereby improving the prediction accuracy and stability of traffic speed on the road network.