Scholarlink

<p>Accurate forecasting of air pollution is critical for public health and environmental management. This paper proposes a hybrid spatiotemporal framework that integrates Neural Controlled Differential Equations (Neural CDE) with Graph Attention Networks (GAT), which is a type of Graph Neural Network (GNN) for 72-hour multi-pollutant forecasting in complex urban settings. The Neural CDE component models continuous-time temporal dynamics through cubic-spline path interpolation, effectively handling irregular and missing sensor data. Meanwhile, the GAT module captures spatial interactions among monitoring stations via an adaptive edge mechanism (edge-MLP), where each edge feature combines the Neural CDE–predicted pollutant states at the source and target stations with physical factors such as inter-station distance, geographic direction, and local wind vectors. Experiments using hourly observations from five stations in Beijing demonstrate that the proposed model achieves an average Mean Absolute Error (MAE) of 13.14, outperforming both CDE-only and GNN-only baselines. The hybrid architecture remains stable and accurate even under incomplete data scenarios, confirming that combining continuous-time neural dynamics with edge-aware graph reasoning yields physically consistent and robust air-quality forecasts.</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

Neural CDE-GAT framework for spatio-temporal air pollution prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月2日

查看全文
  • Sayeh Mirzaei,
  • Erfan Abedinzadeh Torghabeh

摘要

Accurate forecasting of air pollution is critical for public health and environmental management. This paper proposes a hybrid spatiotemporal framework that integrates Neural Controlled Differential Equations (Neural CDE) with Graph Attention Networks (GAT), which is a type of Graph Neural Network (GNN) for 72-hour multi-pollutant forecasting in complex urban settings. The Neural CDE component models continuous-time temporal dynamics through cubic-spline path interpolation, effectively handling irregular and missing sensor data. Meanwhile, the GAT module captures spatial interactions among monitoring stations via an adaptive edge mechanism (edge-MLP), where each edge feature combines the Neural CDE–predicted pollutant states at the source and target stations with physical factors such as inter-station distance, geographic direction, and local wind vectors. Experiments using hourly observations from five stations in Beijing demonstrate that the proposed model achieves an average Mean Absolute Error (MAE) of 13.14, outperforming both CDE-only and GNN-only baselines. The hybrid architecture remains stable and accurate even under incomplete data scenarios, confirming that combining continuous-time neural dynamics with edge-aware graph reasoning yields physically consistent and robust air-quality forecasts.