Scholarlink

<p>Accurate traffic flow forecasting is essential for intelligent transportation systems (ITS). However, existing spatio-temporal forecasting models often suffer from inadequate modeling of dynamic spatial dependencies, insufficient representation of heterogeneous temporal patterns, and excessive computational complexity. To address these issues, this paper proposes a lightweight Dynamic Graph Temporal Decoupling Network (DGTDNet) for multi-step traffic flow forecasting. Specifically, a hybrid graph construction strategy is introduced to combine static road topology with dynamically learned traffic relationships, enabling the model to capture both stable structural dependencies and time-varying spatial interactions. Based on the hybrid graph, a graph convolutional network is employed for spatial feature extraction. In addition, a temporal decoupling module is designed to explicitly separate short-term local fluctuations from long-term temporal dependencies, where a temporal convolution branch captures local dynamics and a Transformer encoder branch models long-range temporal correlations. The two temporal representations are then adaptively integrated through a gated fusion mechanism for final forecasting. Extensive experiments on the public PeMSD4 and PeMSD8 datasets demonstrate the effectiveness of DGTDNet. On the 60-minute forecasting task, DGTDNet achieves Mean Absolute Error (MAE) values of 22.31 and 17.61 on PeMSD4 and PeMSD8, respectively, outperforming MTGNN, the strongest competing baseline, by 3.34% and 3.98%. In addition, DGTDNet consistently achieves the lowest Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) across different forecasting horizons while maintaining competitive computational efficiency. Ablation studies further confirm the effectiveness of the proposed hybrid graph construction, temporal decoupling, and gated fusion strategies. The results indicate that DGTDNet provides an effective and computationally efficient solution for multi-step traffic flow forecasting in ITS.</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

A Lightweight Dynamic Graph Temporal Decoupling Network for Multi-Step Traffic Flow Forecasting

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Chao Liu,
  • Yanguo Huang,
  • Zijie Zhang,
  • Weifeng Liu,
  • Zhenyu Zhong

摘要

Accurate traffic flow forecasting is essential for intelligent transportation systems (ITS). However, existing spatio-temporal forecasting models often suffer from inadequate modeling of dynamic spatial dependencies, insufficient representation of heterogeneous temporal patterns, and excessive computational complexity. To address these issues, this paper proposes a lightweight Dynamic Graph Temporal Decoupling Network (DGTDNet) for multi-step traffic flow forecasting. Specifically, a hybrid graph construction strategy is introduced to combine static road topology with dynamically learned traffic relationships, enabling the model to capture both stable structural dependencies and time-varying spatial interactions. Based on the hybrid graph, a graph convolutional network is employed for spatial feature extraction. In addition, a temporal decoupling module is designed to explicitly separate short-term local fluctuations from long-term temporal dependencies, where a temporal convolution branch captures local dynamics and a Transformer encoder branch models long-range temporal correlations. The two temporal representations are then adaptively integrated through a gated fusion mechanism for final forecasting. Extensive experiments on the public PeMSD4 and PeMSD8 datasets demonstrate the effectiveness of DGTDNet. On the 60-minute forecasting task, DGTDNet achieves Mean Absolute Error (MAE) values of 22.31 and 17.61 on PeMSD4 and PeMSD8, respectively, outperforming MTGNN, the strongest competing baseline, by 3.34% and 3.98%. In addition, DGTDNet consistently achieves the lowest Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) across different forecasting horizons while maintaining competitive computational efficiency. Ablation studies further confirm the effectiveness of the proposed hybrid graph construction, temporal decoupling, and gated fusion strategies. The results indicate that DGTDNet provides an effective and computationally efficient solution for multi-step traffic flow forecasting in ITS.