Scholarlink

<p>Accurate prediction of fine particulate matter (PM<sub>2.5</sub>) remains challenging because atmospheric processes are nonlinear, nonstationary, and influenced by temporal variability, measurement uncertainty, and episodic pollution events. This study develops an interpretable and uncertainty-aware machine-learning (ML) framework for station-level PM<sub>2.5</sub> prediction in Visakhapatnam, a rapidly industrializing coastal urban environment in India characterized by complex meteorological and emission conditions. Multiple ML algorithms and Gradient Boosting loss formulations were comparatively evaluated using a leakage-free chronological validation framework based on daily observations collected during 2018–2024. The proposed framework integrates temporal feature engineering, rolling time-series cross-validation, residual diagnostics, regime-wise evaluation, conformal uncertainty estimation, bootstrap confidence intervals, and SHAP-based interpretability analysis. Predictive performance was assessed using Coefficient of Determination, Root Mean Squared Error, Mean Absolute Error, Mean Absolute Scaled Error, and Mean Absolute Percentage Error metrics across training, validation, and independent testing datasets. Results demonstrate that ensemble-based ML models consistently outperform linear and kernel-based approaches under temporally ordered evaluation conditions. Among the evaluated models, Random Forest achieved the highest overall predictive accuracy on the independent testing dataset, while robustness-oriented Gradient Boosting formulations, particularly the Huber-loss configuration, demonstrated comparatively stable predictive behavior under heterogeneous and noisy atmospheric conditions. Central Pollution Control Board regime-wise evaluation further revealed reduced predictive performance during elevated pollution episodes, emphasizing the importance of pollution-regime-specific evaluation beyond aggregate metrics alone. Overall, the proposed framework provides a computationally efficient, interpretable, and rigorously validated approach for PM<sub>2.5</sub> prediction in complex coastal urban environments, supporting operational air-quality monitoring and environmental decision-support applications.</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 robust and uncertainty-aware machine learning framework for PM2.5 prediction in a coastal urban Indian environment: implications for sustainable air quality management

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月28日

查看全文
  • Salvator Lawrence,
  • Srimuruganandam Bhathmanabhan

摘要

Accurate prediction of fine particulate matter (PM2.5) remains challenging because atmospheric processes are nonlinear, nonstationary, and influenced by temporal variability, measurement uncertainty, and episodic pollution events. This study develops an interpretable and uncertainty-aware machine-learning (ML) framework for station-level PM2.5 prediction in Visakhapatnam, a rapidly industrializing coastal urban environment in India characterized by complex meteorological and emission conditions. Multiple ML algorithms and Gradient Boosting loss formulations were comparatively evaluated using a leakage-free chronological validation framework based on daily observations collected during 2018–2024. The proposed framework integrates temporal feature engineering, rolling time-series cross-validation, residual diagnostics, regime-wise evaluation, conformal uncertainty estimation, bootstrap confidence intervals, and SHAP-based interpretability analysis. Predictive performance was assessed using Coefficient of Determination, Root Mean Squared Error, Mean Absolute Error, Mean Absolute Scaled Error, and Mean Absolute Percentage Error metrics across training, validation, and independent testing datasets. Results demonstrate that ensemble-based ML models consistently outperform linear and kernel-based approaches under temporally ordered evaluation conditions. Among the evaluated models, Random Forest achieved the highest overall predictive accuracy on the independent testing dataset, while robustness-oriented Gradient Boosting formulations, particularly the Huber-loss configuration, demonstrated comparatively stable predictive behavior under heterogeneous and noisy atmospheric conditions. Central Pollution Control Board regime-wise evaluation further revealed reduced predictive performance during elevated pollution episodes, emphasizing the importance of pollution-regime-specific evaluation beyond aggregate metrics alone. Overall, the proposed framework provides a computationally efficient, interpretable, and rigorously validated approach for PM2.5 prediction in complex coastal urban environments, supporting operational air-quality monitoring and environmental decision-support applications.