Scholarlink

<p>The acceleration of urbanization and the growth of motor vehicle numbers have led to increasingly severe traffic congestion problems. Accurate traffic flow prediction is of great significance for optimizing traffic management and alleviating congestion. However, existing methods still have deficiencies in modeling dynamic local correlations, capturing long-term dependencies, and solving the over-smoothing problem. To this end, this paper proposes a novel traffic flow prediction model called Multi-Scale Pooling and Graph Attention Network (MSAGA-Former), which fuses multi-scale adaptive pooling and graph attention mechanisms. The model designs a parallel branch structure, using the Graph Attention Network (GAT) and masked multi-head attention to model dynamic spatial dependencies and local temporal patterns respectively, effectively alleviating the over-smoothing problem in deep networks. Meanwhile, a Multi-scale Aggregation Pyramid Pooling Module (MAPPM) is proposed, combining Global Average Pooling (GAP) and Exponential Adaptive Pooling (AdaPool ) to dynamically extract multi-scale time series features, enhancing the Transformer’s ability to model local correlations. Furthermore, a temporal position encoding module is introduced to explicitly fuse periodic features, improving the model’s learning performance for long-term dependencies. Experimental results show that MSAGA-Former reduces the Mean Absolute Error (MAE) by 7.26% and 11% compared with the best baseline HDCformer on the PeMSD4 and PeMSD8 datasets, respectively. In addition, ablation experiments verify the effectiveness of each key module in the model. Removing the temporal position embedding, MAPPM, or parallel branch structure all lead to significant declines in prediction performance, further confirming the advantages and potential of the proposed method in improving traffic flow prediction accuracy.</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

MSAGA-former: A Novel Multi-scale Adaptive Pooling and Graph Attention Transformer for Traffic Flow Prediction

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Duoduo Zhang,
  • Jiajun Jiang,
  • Wenjing Hu,
  • Shunwei Yao,
  • Lin Peng,
  • Jia Lin

摘要

The acceleration of urbanization and the growth of motor vehicle numbers have led to increasingly severe traffic congestion problems. Accurate traffic flow prediction is of great significance for optimizing traffic management and alleviating congestion. However, existing methods still have deficiencies in modeling dynamic local correlations, capturing long-term dependencies, and solving the over-smoothing problem. To this end, this paper proposes a novel traffic flow prediction model called Multi-Scale Pooling and Graph Attention Network (MSAGA-Former), which fuses multi-scale adaptive pooling and graph attention mechanisms. The model designs a parallel branch structure, using the Graph Attention Network (GAT) and masked multi-head attention to model dynamic spatial dependencies and local temporal patterns respectively, effectively alleviating the over-smoothing problem in deep networks. Meanwhile, a Multi-scale Aggregation Pyramid Pooling Module (MAPPM) is proposed, combining Global Average Pooling (GAP) and Exponential Adaptive Pooling (AdaPool ) to dynamically extract multi-scale time series features, enhancing the Transformer’s ability to model local correlations. Furthermore, a temporal position encoding module is introduced to explicitly fuse periodic features, improving the model’s learning performance for long-term dependencies. Experimental results show that MSAGA-Former reduces the Mean Absolute Error (MAE) by 7.26% and 11% compared with the best baseline HDCformer on the PeMSD4 and PeMSD8 datasets, respectively. In addition, ablation experiments verify the effectiveness of each key module in the model. Removing the temporal position embedding, MAPPM, or parallel branch structure all lead to significant declines in prediction performance, further confirming the advantages and potential of the proposed method in improving traffic flow prediction accuracy.