Scholarlink

<p>Accurate prediction of long-sequence and multi-level air quality indexes across different regions is crucial for pollution control and public health protection. However, current methods face challenges such as insufficient capture of spatiotemporal features, difficulty in achieving global optimization in multi-step forecasting, and a lack of end-to-end internal classification techniques for multi-station data. To address these issues, this study proposes an end-to-end multi-station cooperative prediction model named Gclassify-Mformer. It integrates classification and multi-encoder mechanisms. The model optimizes classification boundaries through error backpropagation, enhances spectral clustering efficiency via graph convolutional network mapping, and strengthens spatiotemporal correlation learning by fusing multiple encoders with graph attention. Experiments on multi-station data from four climatically diverse regions in China demonstrate that the model achieves leading performance in both intra-day (1-step, 12-step, 24-step) and multi-day (48-step, 72-step, 96-step) predictions. For 24-h intra-day forecasting, accuracy improves by 1–3% compared to the best baseline model. Under extreme weather conditions, relevant metrics see improvements of up to 26%. In multi-day long-step predictions, the model outperforms advanced Transformer-based models by 1–3% in accuracy while also delivering superior prediction stability.</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

End-to-end model with internal classification and multi-level feature fusion for collaborative prediction of long-sequence air quality

  • 期刊论文

  • 发表日期：

    2026年3月5日

查看全文
  • Xinjie Shi,
  • Jianzhou Wang,
  • Xiayan Liu,
  • Wenliang Zhang

摘要

Accurate prediction of long-sequence and multi-level air quality indexes across different regions is crucial for pollution control and public health protection. However, current methods face challenges such as insufficient capture of spatiotemporal features, difficulty in achieving global optimization in multi-step forecasting, and a lack of end-to-end internal classification techniques for multi-station data. To address these issues, this study proposes an end-to-end multi-station cooperative prediction model named Gclassify-Mformer. It integrates classification and multi-encoder mechanisms. The model optimizes classification boundaries through error backpropagation, enhances spectral clustering efficiency via graph convolutional network mapping, and strengthens spatiotemporal correlation learning by fusing multiple encoders with graph attention. Experiments on multi-station data from four climatically diverse regions in China demonstrate that the model achieves leading performance in both intra-day (1-step, 12-step, 24-step) and multi-day (48-step, 72-step, 96-step) predictions. For 24-h intra-day forecasting, accuracy improves by 1–3% compared to the best baseline model. Under extreme weather conditions, relevant metrics see improvements of up to 26%. In multi-day long-step predictions, the model outperforms advanced Transformer-based models by 1–3% in accuracy while also delivering superior prediction stability.