Scholarlink

<p>Accurate prediction of air quality holds significant importance for enabling government agencies to formulate effective air pollution prevention and control measures, as well as assisting the public in planning their travel activities more rationally. Given the complex nonlinear coupling relationships between air pollutants and multiple ground-atmosphere factors, this paper proposes a ground-atmosphere multi-factor coupled air quality prediction model (GCN-GAT-LSTM). First, a feature matrix integrating various ground-atmosphere factors is constructed to address the limitation of traditional prediction models that insufficiently account for these factors. Then, a graph convolutional and graph attention network (GCN-GAT) is developed to analyze the topological structure of monitoring stations and extract their spatial features, thereby resolving the issue of spatial correlation among multiple stations. Experiments demonstrate that when the prediction duration is 3&#xa0;h and 6&#xa0;h, the RMSE evaluation metrics reach 31.49 and 38.23 respectively, while the MAE evaluation metrics reach 20.93 and 27.79 respectively, outperforming benchmark models such as SVM, LSTM, GCN, and GAT. Subsequently, to address the issue of temporal correlation among monitoring stations, a ground-atmosphere multi-factor coupled air quality prediction model based on a long short-term memory network (GCN-GAT-LSTM) is constructed. This model extracts temporal features and resolves the long-term dependency problem inherent in time-series air quality prediction. Experiments demonstrate that compared to the GCN-GAT model, the GCN-GAT-LSTM model achieves significant improvements in prediction performance. Specifically, for forecast durations of 3&#xa0;h, 6&#xa0;h, and 12&#xa0;h, the Root Mean Square Error (RMSE) decreases by 2.02, 3.58, and 7.51, respectively, while the Mean Absolute Error (MAE) decreases by 2.4, 4.17, and 8.91, respectively. Finally, ablation experiments were conducted using the GCN-GAT model to analyze the influence of ground-atmosphere interactions. The results indicate that when the land cover type is simple and the terrain is flat, meteorological factors exert a significant influence on the model’s prediction outcomes. Conversely, meteorological factors have a weaker impact compared to land cover type and elevation.</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 Hybrid Neural Network Model for Predicting Air Quality under the Coupling of Ground Atmosphere

  • 期刊论文

  • 发表日期：

    2025年12月8日

查看全文
  • Zhenying Li,
  • Weidong Li,
  • Xuehai Zhang,
  • Shewei Wang,
  • Jinlong Duan,
  • Linyan Bai,
  • Pengchong Lei,
  • Wentao Chen,
  • Xinguang Li

摘要

Accurate prediction of air quality holds significant importance for enabling government agencies to formulate effective air pollution prevention and control measures, as well as assisting the public in planning their travel activities more rationally. Given the complex nonlinear coupling relationships between air pollutants and multiple ground-atmosphere factors, this paper proposes a ground-atmosphere multi-factor coupled air quality prediction model (GCN-GAT-LSTM). First, a feature matrix integrating various ground-atmosphere factors is constructed to address the limitation of traditional prediction models that insufficiently account for these factors. Then, a graph convolutional and graph attention network (GCN-GAT) is developed to analyze the topological structure of monitoring stations and extract their spatial features, thereby resolving the issue of spatial correlation among multiple stations. Experiments demonstrate that when the prediction duration is 3 h and 6 h, the RMSE evaluation metrics reach 31.49 and 38.23 respectively, while the MAE evaluation metrics reach 20.93 and 27.79 respectively, outperforming benchmark models such as SVM, LSTM, GCN, and GAT. Subsequently, to address the issue of temporal correlation among monitoring stations, a ground-atmosphere multi-factor coupled air quality prediction model based on a long short-term memory network (GCN-GAT-LSTM) is constructed. This model extracts temporal features and resolves the long-term dependency problem inherent in time-series air quality prediction. Experiments demonstrate that compared to the GCN-GAT model, the GCN-GAT-LSTM model achieves significant improvements in prediction performance. Specifically, for forecast durations of 3 h, 6 h, and 12 h, the Root Mean Square Error (RMSE) decreases by 2.02, 3.58, and 7.51, respectively, while the Mean Absolute Error (MAE) decreases by 2.4, 4.17, and 8.91, respectively. Finally, ablation experiments were conducted using the GCN-GAT model to analyze the influence of ground-atmosphere interactions. The results indicate that when the land cover type is simple and the terrain is flat, meteorological factors exert a significant influence on the model’s prediction outcomes. Conversely, meteorological factors have a weaker impact compared to land cover type and elevation.