Scholarlink

<p>The persistent trend of urbanization is catalyzing a consistent increment in the volume of data gathered via sophisticated environmental monitoring frameworks, which is of great significance for air quality prediction. Nevertheless, the prediction models predominantly employed in air quality forecasting endeavors predominantly adhere to conventional temporal and spatial dimensionality techniques. Especially in scenarios where pollution sources are unevenly distributed, the number and types of monitoring stations have increased substantially, leading to substantial growth in the intricacy and multi-dimensionality of air quality data. Moreover, the lack of exploration into research models that incorporate the interplay between various dimensions has led to suboptimal predictive outcomes in complex environmental contexts. This paper proposes a novel framework termed the Triple Cross-Attention and Self-Loop Normalized Graph Attention Network (TC-SLNGAT). This framework operates seamlessly across spatial–temporal and feature dimensions, effectively tackling both the inherent complexity of spatio-temporal dependencies in air quality prediction and the intricate cross-dimensional dependencies among these three dimensions. For an input tensor, the cross-attention mechanism establishes inter-dimensional dependencies through rotation operations and residual transformations. By constructing cross-informational perspectives for spatial–temporal and feature dimensions, TC-SLNGAT not only captures the complex spatio-temporal dependencies in air quality prediction but also adeptly handles the cross-dimensional dependencies that exist among these three dimensions. The TC-SLNGAT is validated on real-world Shanghai and Nanjing datasets, benchmarked against mainstream methods. The outcomes highlight 13.25% Mean Absolute Error (MAE) and 14.26% Root Mean Square Error (RMSE) improvements in Shanghai, and 9.61% MAE,10.98% RMSE enhancements in Nanjing versus baselines, validating the approach’s effectiveness.</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 short-term prediction approach for air quality based on Triple cross-attention and self-loop normalized Graph Attention Network

  • 期刊论文

  • 发表日期：

    2025年4月17日

查看全文
  • Chaoli Tang,
  • Yi Hu,
  • Yuanyuan Wei,
  • Ziyu Wang,
  • Zhiwei Zhao

摘要

The persistent trend of urbanization is catalyzing a consistent increment in the volume of data gathered via sophisticated environmental monitoring frameworks, which is of great significance for air quality prediction. Nevertheless, the prediction models predominantly employed in air quality forecasting endeavors predominantly adhere to conventional temporal and spatial dimensionality techniques. Especially in scenarios where pollution sources are unevenly distributed, the number and types of monitoring stations have increased substantially, leading to substantial growth in the intricacy and multi-dimensionality of air quality data. Moreover, the lack of exploration into research models that incorporate the interplay between various dimensions has led to suboptimal predictive outcomes in complex environmental contexts. This paper proposes a novel framework termed the Triple Cross-Attention and Self-Loop Normalized Graph Attention Network (TC-SLNGAT). This framework operates seamlessly across spatial–temporal and feature dimensions, effectively tackling both the inherent complexity of spatio-temporal dependencies in air quality prediction and the intricate cross-dimensional dependencies among these three dimensions. For an input tensor, the cross-attention mechanism establishes inter-dimensional dependencies through rotation operations and residual transformations. By constructing cross-informational perspectives for spatial–temporal and feature dimensions, TC-SLNGAT not only captures the complex spatio-temporal dependencies in air quality prediction but also adeptly handles the cross-dimensional dependencies that exist among these three dimensions. The TC-SLNGAT is validated on real-world Shanghai and Nanjing datasets, benchmarked against mainstream methods. The outcomes highlight 13.25% Mean Absolute Error (MAE) and 14.26% Root Mean Square Error (RMSE) improvements in Shanghai, and 9.61% MAE,10.98% RMSE enhancements in Nanjing versus baselines, validating the approach’s effectiveness.