Scholarlink

<p>Rapid urbanization and industrialization have intensified air pollution in smart cities, necessitating accurate, real-time Air Quality Index (AQI) prediction systems. Existing federated learning approaches for AQI prediction share three critical limitations: reliance on homogeneous client assumptions that exclude resource-constrained sensors with partial feature availability, dependence on simulated rather than real heterogeneous data sources, and the absence of formal cryptographic privacy mechanisms with empirically quantified accuracy tradeoffs. This paper presents FedLSTM-AQI, a federated deep learning framework that directly addresses each of these gaps. Three genuinely heterogeneous clients are integrated reflecting real smart city deployment conditions: a government Central Pollution Control Board (CPCB) monitoring station (2017–2023), an outdoor IoT sensor, and an indoor sensor deployed at Dr. B.R. Ambedkar National Institute of Technology (NIT) Jalandhar campus. A structural feature-alignment strategy based on zero-padding, adopted as a practical engineering approximation, enables participation of feature-incomplete clients without modifying the global model architecture. Both Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) architectures are evaluated within the federated framework. Paillier homomorphic encryption is applied to the output layer to provide a computationally scoped cryptographic privacy mechanism, with the privacy–accuracy tradeoff empirically quantified across baseline, federated, and encrypted configurations. The BiLSTM model achieves <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2 = 0.9819\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.9819</mn> </mrow> </math></EquationSource> </InlineEquation> on CPCB benchmark data, with negligible degradation to <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2 = 0.9790\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.9790</mn> </mrow> </math></EquationSource> </InlineEquation> under federated training on real sensor data, and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(R^2 = 0.9629\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <msup> <mi>R</mi> <mn>2</mn> </msup> <mo>=</mo> <mn>0.9629</mn> </mrow> </math></EquationSource> </InlineEquation> retained under homomorphic encryption.</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

FedLSTM-AQI: a federated deep learning framework for air quality index prediction

  • 期刊论文

  • 发表日期：

    2026年7月17日

查看全文
  • Jaspal Kaur Saini,
  • Manpreet Singh,
  • Divya Bansal

摘要

Rapid urbanization and industrialization have intensified air pollution in smart cities, necessitating accurate, real-time Air Quality Index (AQI) prediction systems. Existing federated learning approaches for AQI prediction share three critical limitations: reliance on homogeneous client assumptions that exclude resource-constrained sensors with partial feature availability, dependence on simulated rather than real heterogeneous data sources, and the absence of formal cryptographic privacy mechanisms with empirically quantified accuracy tradeoffs. This paper presents FedLSTM-AQI, a federated deep learning framework that directly addresses each of these gaps. Three genuinely heterogeneous clients are integrated reflecting real smart city deployment conditions: a government Central Pollution Control Board (CPCB) monitoring station (2017–2023), an outdoor IoT sensor, and an indoor sensor deployed at Dr. B.R. Ambedkar National Institute of Technology (NIT) Jalandhar campus. A structural feature-alignment strategy based on zero-padding, adopted as a practical engineering approximation, enables participation of feature-incomplete clients without modifying the global model architecture. Both Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) architectures are evaluated within the federated framework. Paillier homomorphic encryption is applied to the output layer to provide a computationally scoped cryptographic privacy mechanism, with the privacy–accuracy tradeoff empirically quantified across baseline, federated, and encrypted configurations. The BiLSTM model achieves \(R^2 = 0.9819\) R 2 = 0.9819 on CPCB benchmark data, with negligible degradation to \(R^2 = 0.9790\) R 2 = 0.9790 under federated training on real sensor data, and \(R^2 = 0.9629\) R 2 = 0.9629 retained under homomorphic encryption.