Scholarlink

<p>Accurate forecasting of air quality is critical, as PM<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(_{2.5}\)</EquationSource> </InlineEquation> pollution poses serious risks to environmental sustainability and public health. This study proposes a novel hybrid deep ensemble framework integrating CNN, BiLSTM, and GRU within a stacked architecture, with a Gradient Boosting Machine (GBM) serving as the meta-learner. Using daily PM<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(_{2.5}\)</EquationSource> </InlineEquation> data from Delhi, India, the framework is benchmarked against classical and deep learning models. Results demonstrate that the hybrid ensemble consistently outperforms baselines, achieving MAE of 8.09, RMSE of 10.70, and R<InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(^{2}\)</EquationSource> </InlineEquation> of 0.96, representing a 27.80% improvement in predictive correlation over the cutting-edge hybrid BiLSTM–GRU model. These findings highlight the effectiveness of ensemble learning in modeling complex pollution dynamics and underscore the potential of the proposed framework as a reliable decision-support tool for policymakers and urban planners.</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

Advancing PM\(_{2.5}\) Forecasting with Hybrid Deep Learning Ensembles: Application to Delhi’s Air Quality

  • 期刊论文

  • 发表日期：

    2026年7月23日

查看全文
  • Aditya Kumar,
  • Ravi Patel,
  • Niharika Koch,
  • Jainath Yadav

摘要

Accurate forecasting of air quality is critical, as PM \(_{2.5}\) pollution poses serious risks to environmental sustainability and public health. This study proposes a novel hybrid deep ensemble framework integrating CNN, BiLSTM, and GRU within a stacked architecture, with a Gradient Boosting Machine (GBM) serving as the meta-learner. Using daily PM \(_{2.5}\) data from Delhi, India, the framework is benchmarked against classical and deep learning models. Results demonstrate that the hybrid ensemble consistently outperforms baselines, achieving MAE of 8.09, RMSE of 10.70, and R \(^{2}\) of 0.96, representing a 27.80% improvement in predictive correlation over the cutting-edge hybrid BiLSTM–GRU model. These findings highlight the effectiveness of ensemble learning in modeling complex pollution dynamics and underscore the potential of the proposed framework as a reliable decision-support tool for policymakers and urban planners.