机构地区:[1]武汉大学测绘遥感信息工程国家重点实验室,武汉430079 [2]武汉大学社会地理计算联合研究中心,武汉430079
出 处:《地球信息科学学报》2025年第3期653-667,共15页Journal of Geo-information Science
基 金:国家重点研发计划项目(2021YFB3900903);国家自然科学基金项目(42271473)。
摘 要:【目的】准确、可靠的交通状态预测信息是智能交通系统(ITS)中许多应用的基础。然而,城市道路网络连接关系复杂,且存在较大的空间依赖性建模挑战。传统的图卷积神经网络(GCN)广泛应用于交通预测中,但其并未专门为交通问题设计,无法充分考虑交通网络的特性,如路段的行驶方向、转向规则等。因此,本研究旨在提出一种新的图卷积网络模型,能够有效建模城市交通路网中的转向关系,从而提升交通预测精度。【方法】研究从交通网络的特性出发,考虑到中心路段与其邻居路段的转向关系和相对位置构建了新型的图卷积网络,称为转向图卷积神经网络(Turn-based Graph Convolution Neural Network,TurnGCN)。TurnGCN方法将城市路网建模为异构图,图中的边代表路段间的转向关系。在计算过程中,TurnGCN通过引入转向表来标记邻居节点,并将节点特征映射为欧氏空间的特征栅格后,利用卷积神经网络(CNN)对节点特征进行融合。【结果】在2个现实世界交通数据集上进行了模型的对比实验:韩国首尔的Urban-150数据集和中国上海的SHSpeed数据集。实验结果表明,在多个评估指标上,TurnGCN的预测精度均达到最高,且参数量小于基于空间注意力机制的GCN变体。【结论】TurnGCN为城市路网交通预测提供了一种有效的解决方案,显式地建模了基于转向的空间关系。它有效克服了传统GCN和注意力机制模型的局限,显著提高了预测性能,同时参数共享的机制保持了模型的计算效率。这些优势突出了TurnGCN在智能交通系统中的实际应用潜力,尤其在交通流量优化、拥堵管理和智能导航系统中具有广泛的应用前景。[Objectives]Accurate and reliable traffic state prediction is essential for various applications in Intelligent Transportation Systems(ITS).However,the complexity of urban road networks makes it challenging to effectively model spatial dependencies between road segments,posing a significant obstacle to urban traffic forecasting.Traditional Graph Convolutional Networks(GCNs)are widely used for traffic prediction but fail to account for the unique characteristics of traffic networks,such as driving directions,turning rules,and varying spatial dependencies.This study aims to address these challenges by proposing a novel graph convolutional network model,the Turn-based Graph Convolutional Neural Network(TurnGCN),which better captures the complex spatial relationships in urban traffic networks.[Methods]TurnGCN models the urban road network as a heterogeneous graph,where edges represent turning relationships between road segments.Unlike traditional GCNs that rely on static adjacency matrices,TurnGCN introduces a turning table to label neighboring nodes and map their features into a structured Euclidean feature grid.A Convolutional Neural Network(CNN)is then applied to this grid to aggregate and fuse the spatial features of neighboring nodes.This approach allows TurnGCN to model the heterogeneity of turning relationships and learn their varying impacts on the central road segment.Additionally,the parameter-sharing nature of CNNs ensures that TurnGCN performs efficiently with relatively fewer trainable parameters.[Results]To validate the effectiveness of TurnGCN,extensive experiments were conducted on two real-world traffic datasets:Urban-150 from Seoul,South Korea,and SHSpeed from Shanghai,China.These datasets vary in sampling density and temporal resolution,presenting diverse evaluation challenges.The results demonstrate that TurnGCN consistently outperforms traditional GCNs and GCN variants enhanced with spatial attention mechanisms across multiple evaluation metrics.Specifically,TurnGCN excels in capturing heterogene
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...
正在载入数据...