Scholarlink

<p>Rapid urbanization and industrial growth in developing countries have intensified environmental degradation, prominently exemplified by increasing concentrations of ambient particulate matter (PM<sub>2.5</sub>), significantly affecting public health and the economy. This study focuses on predicting daily PM<sub>2.5</sub> levels in Delhi, India, a city characterized by complex air quality dynamics due to diverse anthropogenic and meteorological influences. To capture these complexities, we leveraged data from 39 monitoring stations (2019 – 2023) and developed a multi-model framework, employing statistical approaches (SARIMAX), machine learning algorithms (Random Forest, Support Vector Machines), and deep learning models (Artificial Neural Networks, Long Short-Term Memory Networks). The framework uniquely combines station-specific hyperparameter optimization, comprehensive exogenous variables (co-pollutants: PM₁₀, NO₂, SO₂, O₃, CO; and meteorological parameters), and Fourier-transformed functions which explicitly capture the multi-scale seasonal variations. Modeling results reveal ANN as the top performer (testing R<sup>2</sup> = 0.81–0.98, RMSE = 10.75–31.96 µg/m<sup>3</sup>, MAE = 7.71–19.42 µg/m<sup>3</sup>), followed closely by Bi-LSTM (R<sup>2</sup> = 0.79–0.96, RMSE = 13.99–32.45 µg/m<sup>3</sup>, MAE = 8.05–18.92 µg/m<sup>3</sup>). RF demonstrated robust intermediate accuracy (R<sup>2</sup> = 0.79–0.95, RMSE = 15.33–34.87 µg/m<sup>3</sup>, MAE = 9.84–24.13 µg/m<sup>3</sup>), outperforming SARIMAX and SVM models. Fourier terms enhanced prediction stability by capturing seasonal dynamics, while station-specific tuning improved localized accuracy. These robust predictions are crucial for timely public health advisories and evidence-based policymaking, ultimately aiming to mitigate health risks and facilitate sustainable urban environmental management in one of the world’s most polluted cities.</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

Fourier-Enhanced Deep Learning and Machine Learning Models for Predicting Multi-Scale PM2.5 Dynamics in Megacities: A Case Study of Delhi

  • 期刊论文

  • 发表日期：

    2025年9月5日

查看全文
  • Divyansh Sharma,
  • Sapan Thapar,
  • Adil Masood,
  • Kamna Sachdeva

摘要

Rapid urbanization and industrial growth in developing countries have intensified environmental degradation, prominently exemplified by increasing concentrations of ambient particulate matter (PM2.5), significantly affecting public health and the economy. This study focuses on predicting daily PM2.5 levels in Delhi, India, a city characterized by complex air quality dynamics due to diverse anthropogenic and meteorological influences. To capture these complexities, we leveraged data from 39 monitoring stations (2019 – 2023) and developed a multi-model framework, employing statistical approaches (SARIMAX), machine learning algorithms (Random Forest, Support Vector Machines), and deep learning models (Artificial Neural Networks, Long Short-Term Memory Networks). The framework uniquely combines station-specific hyperparameter optimization, comprehensive exogenous variables (co-pollutants: PM₁₀, NO₂, SO₂, O₃, CO; and meteorological parameters), and Fourier-transformed functions which explicitly capture the multi-scale seasonal variations. Modeling results reveal ANN as the top performer (testing R2 = 0.81–0.98, RMSE = 10.75–31.96 µg/m3, MAE = 7.71–19.42 µg/m3), followed closely by Bi-LSTM (R2 = 0.79–0.96, RMSE = 13.99–32.45 µg/m3, MAE = 8.05–18.92 µg/m3). RF demonstrated robust intermediate accuracy (R2 = 0.79–0.95, RMSE = 15.33–34.87 µg/m3, MAE = 9.84–24.13 µg/m3), outperforming SARIMAX and SVM models. Fourier terms enhanced prediction stability by capturing seasonal dynamics, while station-specific tuning improved localized accuracy. These robust predictions are crucial for timely public health advisories and evidence-based policymaking, ultimately aiming to mitigate health risks and facilitate sustainable urban environmental management in one of the world’s most polluted cities.