| 摘 要: 针对交通流量预测中复杂时序建模能力不足与空间全局依赖建模缺失的问题，提出了一种基于时间序列大模型(Large Time Series Model, LTSM)和图卷积网络(Convolutional Network, GCN)的交通流量预测方法。首先，基于路网结构和流量相似性构建地理-语义双空间图；其次，利用LTSM建模交通节点的时序依赖，再通过GCN提取局部物理与全局语义特征；最后，融合两种特征并经全连接层映射得到预测值。实验结果表明，所提方法的平均预测误差较最优基线降低了4.3%，充分验证了该方法的有效性。 |
| 关键词: 交通流量预测 时间序列大模型 图卷积网络 智能交通系统 |
|
中图分类号:
文献标识码:
|
|
| Traffic Flow Prediction Based on Large Time Series Model and Graph Convolutional Network |
|
LiLu1, OuWenxiang2
|
1.Nanjing Jiangbei Smart Transportation Co., Ltd.;2.Nanjing Institute of Software Technology
|
| Abstract: To address the insufficient complex temporal modeling and the lack of spatial global dependency modeling in traffic flow prediction, a method based on Large Time Series Model (LTSM) and Graph Convolutional Network (GCN) is proposed. First, a geo-semantic dual-space graph is constructed based on road network topology and flow similarity. Second, LTSM is employed to model the temporal dependencies of traffic nodes, and GCN is then utilized to extract local physical and global semantic features. Finally, the two features are fused and mapped to the prediction space via a fully-connected layer to obtain the predicted values. Experimental results demonstrate that the proposed method reduces the average prediction error by 4.3% compared to the best baseline, fully validating its effectiveness. |
| Keywords: traffic flow prediction large time series model graph convolutional network intelligent transportation system |
