Scholarlink

<p>Existing convolutional recurrent neural networks, such as ConvLSTM and ConvGRU, are designed to model spatiotemporal features by encoding sequence information into three-dimensional tensors. While these architectures partially alleviate the limitations of conventional recurrent neural networks (RNNs) in capturing spatial dependencies for traffic flow prediction tasks, they still suffer from certain limitations. Specifically, the hidden spatiotemporal states in these models propagate unidirectionally and remain independent across temporal and spatial dimensions, hindering the effective fusion of hierarchical spatiotemporal features within recurrent network units. To address these challenges, we propose an Attention-Augmented Adaptive Graph Convolutional Recurrent Network (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\hbox {A}^3\)</EquationSource> </InlineEquation>GCRN) for multi-step traffic flow prediction. First, we design a novel encoder component, ST-GRU, by integrating an adaptive graph convolutional operator with a gated recurrent unit (GRU). In the encoding network, we employ a dual hidden-states transition mechanism to achieve unified modeling of temporal features and hierarchical spatial dependencies. Additionally, in the decoding process, we introduce an exponential-decay temporal attention mechanism to model both content and temporal dependencies of dynamic context vectors at each decoding step, thereby enhancing sequence-to-sequence temporal dependency learning. Experimental results demonstrate that <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\hbox {A}^3\)</EquationSource> </InlineEquation>GCRN, based on ST-GRU units, significantly outperforms existing baseline models in terms of both prediction accuracy and stability across multiple traffic flow datasets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An attention-augmented adaptive graph convolutional recurrent network for multi-step traffic flow prediction

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Yang Chen,
  • Zikang Dai,
  • Liming Jiang,
  • Huanyu Wang,
  • Shaomiao Chen

摘要

Existing convolutional recurrent neural networks, such as ConvLSTM and ConvGRU, are designed to model spatiotemporal features by encoding sequence information into three-dimensional tensors. While these architectures partially alleviate the limitations of conventional recurrent neural networks (RNNs) in capturing spatial dependencies for traffic flow prediction tasks, they still suffer from certain limitations. Specifically, the hidden spatiotemporal states in these models propagate unidirectionally and remain independent across temporal and spatial dimensions, hindering the effective fusion of hierarchical spatiotemporal features within recurrent network units. To address these challenges, we propose an Attention-Augmented Adaptive Graph Convolutional Recurrent Network ( \(\hbox {A}^3\) GCRN) for multi-step traffic flow prediction. First, we design a novel encoder component, ST-GRU, by integrating an adaptive graph convolutional operator with a gated recurrent unit (GRU). In the encoding network, we employ a dual hidden-states transition mechanism to achieve unified modeling of temporal features and hierarchical spatial dependencies. Additionally, in the decoding process, we introduce an exponential-decay temporal attention mechanism to model both content and temporal dependencies of dynamic context vectors at each decoding step, thereby enhancing sequence-to-sequence temporal dependency learning. Experimental results demonstrate that \(\hbox {A}^3\) GCRN, based on ST-GRU units, significantly outperforms existing baseline models in terms of both prediction accuracy and stability across multiple traffic flow datasets.