多通道自适应特征融合的城市路网交通流量预测  

Multi-Channel Adaptive Feature Fusion for Urban Road Network Traffic Prediction

作　　者：马蕴一 许明 金海波 MAYunyi;XU Ming;JIN Haibo(Software College,Liaoning Technical University,Huludao,Liaoning 125105,China)

机构地区：[1]辽宁工程技术大学软件学院,辽宁葫芦岛125105

出　　处：《计算机工程与应用》2025年第7期334-341,共8页Computer Engineering and Applications

基　　金：国家自然科学基金(62173171)。

摘　　要：精准预测城市路网交通流量对交通管控至关重要。然而,由于城市路网的时空相关性非常复杂,精准交通流量预测极具挑战。为了更好捕捉路网时空相关性,提出了基于车辆轨迹数据的多通道自适应特征融合城市路网交通流量预测模型(multi-channel adaptive feature fusion network,MCAFF-Net)。提出旅行路径编解码器捕获路网的路径级空间关联。根据路段特征向量构建路网特征空间相似度图,并提出拓扑增强的特征向量相关性网络,以建模路段特征空间相关性。提出了长短期依赖图神经网络,以增大图神经网络的感受野,同时捕捉路网的局部和全局空间相关性。采用自适应学习方式进行特征融合,捕捉路网丰富的动态空间关联。设计了稀疏注意力时间相关性模块,捕捉路网的时间相关性,并降低传统注意力模型的计算复杂度。实验结果表明,与现有先进的基线算法相比,MCAFF-Net预测效果最佳。具体而言,在Sumo-SY和Taxi-BJ数据集的第一个时间步预测中,RMSE指标分别降低了12.89%和5.4%,MAE指标分别降低了13.92%和3.3%。此外,通过与该模型的六种变体进行消融实验,验证了该模型各组件的有效性。Accurate prediction of urban road traffic flow is crucial for traffic management.However,precise forecasting faces significant challenges due to the complex spatio-temporal correlations inherent in urban road networks.To better capture the correlations,it proposes the multi-channel adaptive feature fusion network(MCAFF-Net),which is based on vehicle trajectory data.Firstly,it introduces a travel path encoder-decoder to capture path-level spatial correlations in the road network.Secondly,it constructs a road network feature space similarity graph based on segment feature vectors and proposes a topologically enhanced feature vector correlation network to capture segment feature space correlations.Next,it introduces a long short-term dependency graph neural network to enlarge the receptive field of the graph neural network while capturing both local and global spatial correlations in the road network.Subsequently,adaptive learning is employed for feature fusion to capture the dynamic spatial correlations in the road network.Finally,it designs a sparse attention temporal correlation module to capture temporal correlations in the road network while reducing the computational complexity of traditional attention models.Experimental results demonstrate that MCAFF-Net outperforms existing state-of-the-art baseline algorithms.Specifically,in the initial time step prediction on the Sumo-SY and Taxi-BJ datasets,the RMSE indicators are reduced by 12.89%and 5.4%,respectively,and the MAE indicators are reduced by 13.92%and 3.3%,respectively.Furthermore,ablation experiments conducted with six variants of the model validate the effectiveness of each component.

关 键 词：交通流量预测 图神经网络 稀疏注意力机制 特征融合 车辆轨迹挖掘 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]