Scholarlink

Accurate forecasting of PM10 concentration is crucial for air quality management and the reliable operation of renewable energy systems. However, the complex spatiotemporal dependencies and heterogeneous influencing factors pose significant challenges. This paper proposes a novel framework, RTAF-STGNN-MoE, that integrates Residual Attention Fusion (RTAF), a Spatiotemporal Graph Neural Network (STGNN), and a Mixture-of-Experts (MoE) module. RTAF enhances multi-scale temporal feature extraction, while the dynamically constructed multimodal graph incorporates geographic proximity, meteorological similarity, and historical pattern similarity for adaptive spatial modeling. The sparse MoE module further enables dynamic expert selection for different regions and scenarios, improving generalization and robustness. Experiments on a real-world multi-station PM10 dataset demonstrate that the proposed model outperforms state-of-the-art methods in multiple metrics, achieving the highest R2 score. Ablation studies validate the effectiveness of each module. The proposed model provides a scalable and accurate solution for complex air pollution forecasting tasks.

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

Spatiotemporal PM10 Concentration Forecasting via Residual Attention Fusion and Mixture-of-Experts Enhanced Graph Neural Network

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Guolin Zhang,
  • Yongsheng Wang,
  • Delong Zhang

摘要

Accurate forecasting of PM10 concentration is crucial for air quality management and the reliable operation of renewable energy systems. However, the complex spatiotemporal dependencies and heterogeneous influencing factors pose significant challenges. This paper proposes a novel framework, RTAF-STGNN-MoE, that integrates Residual Attention Fusion (RTAF), a Spatiotemporal Graph Neural Network (STGNN), and a Mixture-of-Experts (MoE) module. RTAF enhances multi-scale temporal feature extraction, while the dynamically constructed multimodal graph incorporates geographic proximity, meteorological similarity, and historical pattern similarity for adaptive spatial modeling. The sparse MoE module further enables dynamic expert selection for different regions and scenarios, improving generalization and robustness. Experiments on a real-world multi-station PM10 dataset demonstrate that the proposed model outperforms state-of-the-art methods in multiple metrics, achieving the highest R2 score. Ablation studies validate the effectiveness of each module. The proposed model provides a scalable and accurate solution for complex air pollution forecasting tasks.