基于时间感知注意力与拥塞驱动图卷积的交通流量预测
Traffic Flow Prediction Model Based on Temporal-aware Attention and Congestion-driven Graph Convolutional Network
王小军 1王兴 2林羽 1金彪 2熊金波2
作者信息
- 1. 福建师范大学计算机与网络空间安全学院,福建 福州 350117
- 2. 福建师范大学计算机与网络空间安全学院,福建 福州 350117;福建省大数据分析与应用工程研究中心,福建 福州 350117
- 折叠
摘要
现有的交通流量预测模型通常基于时空图神经网络(GNN)提取交通流量数据的时空特征.然而,GNN通常基于路网连通性、路网中传感器之间的距离构建图结构,忽略了路段的拥塞信息.为了充分捕捉交通流量数据的动态时空依赖关系,提出一种基于时间感知注意力与拥塞驱动图卷积的交通流量预测模型TCGCN.在空间维度上,TCGCN利用两种衡量交通拥塞程度的度量指标构建图结构,挖掘路段之间的动态空间依赖关系,在时间维度上,TCGCN设计了一个时间门控因果卷积多头自注意力机制增强模型的时间感知能力和局部趋势感知能力.最后,在两个真实交通数据集上的实验结果表明,与最优基线模型相比,TCGCN在MAE、MAPE和RMSE这 3 评价指标上分别平均降低了 5.22%、5.05%和 2.90%.
Abstract
Current traffic flow prediction models typically rely on spatiotemporal graph neural networks(GNN)to extract spatiotemporal features.However,GNN generally constructs graph structures based on road network connectivity and the distance between sensors,often overlooking congestion information in road segments.To fully capture the dynamic spatiotemporal dependencies in traffic flow data,this paper proposes a spatiotemporal traffic flow prediction model based on Temporal-Aware Attention and Congestion-Driven Graph Convolutional Network(TCGCN).In the spatial dimension,TCGCN employs two metrics to measure traffic congestion,which are utilized to construct graph structures that reveal dynamic spatial dependencies between road segments.In the temporal dimension,TCGCN designs a time-gated causal convolutional multi-head self-attention mechanism to enhance the model's temporal awareness and local trend perception capabilities.Experimental results on two real-world traffic datasets demonstrate that TCGCN outperforms the optimal baseline model,achieving average reductions of 5.22%,5.06%,and 2.90%in the evaluation metrics of MAE,MAPE,and RMSE,respectively.
关键词
交通流量预测/交通拥塞/注意力机制/时空图神经网络Key words
traffic flow prediction/traffic congestion/attention mechanism/GNN引用本文复制引用
出版年
2025
国家科技期刊平台
NSTL
万方数据