Scholarlink

<p>Air quality is a significant health concern for the twenty-first century, and the need for predictive models cannot be overstated. However, the recent surge in the development of deep learning has led to a fragmented literature, with algorithmic advancements being treated as distinct families of models rather than the continuous mathematical progression they actually are. In this systematic review, a total of 123 articles were compiled from the years 2018 to 2025. Instead of the conventional descriptive taxonomies, air quality prediction is formulated as a spatio-temporal operator learning problem. CNN, LSTM, GNN, and Transformers are all expressed as parameterized approximations of the underlying advection-diffusion-reaction physical equations. From the normalized cross-study meta-analysis, conventional hybrid CNN-LSTM architectures were observed to perform as localized discrete operators, which, although robust, were observed to accumulate autoregressive error during extended forecasts. In contrast, recent architectures such as Graph Neural Networks (GNN) and Transformers were observed to perform as non-Euclidean manifold operators and continuous integral operators, respectively, and were observed to have superior robustness and up to 8% relative error reduction during extended 72-hour forecasts. Additionally, the integration of Explainable AI (XAI) has been critically evaluated, progressing from conventional post-hoc scalar attribution to intrinsic operator transparency. The path forward for the field of predictive modeling appears to be the integration of data-driven Neural Operators with physics-based Chemical Transport Models (CTMs), which would ensure physically plausible and globally generalizable prediction systems.</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

Deep Learning for Air Pollution Prediction: A Systematic Review of CNN, LSTM, GNN, and Transformer Architectures

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Hiteshri Yagnik,
  • Rajeev Kumar Gupta,
  • Aum Pandya

摘要

Air quality is a significant health concern for the twenty-first century, and the need for predictive models cannot be overstated. However, the recent surge in the development of deep learning has led to a fragmented literature, with algorithmic advancements being treated as distinct families of models rather than the continuous mathematical progression they actually are. In this systematic review, a total of 123 articles were compiled from the years 2018 to 2025. Instead of the conventional descriptive taxonomies, air quality prediction is formulated as a spatio-temporal operator learning problem. CNN, LSTM, GNN, and Transformers are all expressed as parameterized approximations of the underlying advection-diffusion-reaction physical equations. From the normalized cross-study meta-analysis, conventional hybrid CNN-LSTM architectures were observed to perform as localized discrete operators, which, although robust, were observed to accumulate autoregressive error during extended forecasts. In contrast, recent architectures such as Graph Neural Networks (GNN) and Transformers were observed to perform as non-Euclidean manifold operators and continuous integral operators, respectively, and were observed to have superior robustness and up to 8% relative error reduction during extended 72-hour forecasts. Additionally, the integration of Explainable AI (XAI) has been critically evaluated, progressing from conventional post-hoc scalar attribution to intrinsic operator transparency. The path forward for the field of predictive modeling appears to be the integration of data-driven Neural Operators with physics-based Chemical Transport Models (CTMs), which would ensure physically plausible and globally generalizable prediction systems.