The increasing challenge of air pollution in cities requires smart methods to make proper predictions and manage the problem. Although machine learning and deep learning models have contributed greatly to weather and pollution forecasting, the main issue is the real-time flexibility, and scalability in the varying atmospheric conditions. This paper introduces a weighted Voting ensemble model that combines Gradient Boosting (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>4), CatBoost (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>3), XGBoost (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>2) and LightGBM (<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\times\)</EquationSource> </InlineEquation>1) to improve the accuracy of Air Quality Index (AQI) forecasting. The full preprocessing (complete-case deletion, which retains extremes) and optimization of hyperparameters (GridSearchCV/Optuna, 5-fold CV) were used to enhance the robustness and generalizability of the model. The Taiwan Air Quality Dataset (2016–2024, <InlineEquation ID="IEq5"> <EquationSource Format="TEX">\(n=4.6M\)</EquationSource> </InlineEquation> hourly records from 74 stations), 6 major pollutants (PM2.5, PM10, <InlineEquation ID="IEq6"> <EquationSource Format="TEX">\(\hbox {NO}_2\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq7"> <EquationSource Format="TEX">\(\hbox {O}_3\)</EquationSource> </InlineEquation>, <InlineEquation ID="IEq8"> <EquationSource Format="TEX">\(\hbox {SO}_2\)</EquationSource> </InlineEquation>, CO), meteorological parameters (wind speed/direction), and 8-h averages) is used to model the data (spatial/temporal IDs are excluded, to allow deployment to a single station). Experimental validation of 60/16/24 splits (random + temporal validation) shows that the ensemble has validation MSE 0.6553 (<InlineEquation ID="IEq9"> <EquationSource Format="TEX">\(\hbox {R}^2\)</EquationSource> </InlineEquation> 0.9969), which beats 15 baselines including the deep learning (LSTM MSE 45.4), but has temporal robustness (<InlineEquation ID="IEq10"> <EquationSource Format="TEX">\(\Delta\)</EquationSource> </InlineEquation> <InlineEquation ID="IEq11"> <EquationSource Format="TEX">\(\hbox {R}^2\)</EquationSource> </InlineEquation>= − 0.0037). Moreover, SHAP is implemented to offer explainability, as it gives more insights into the contribution of features in predicting AQI. The results indicate the promise of interpretable ensemble learning systems to underpin sustainable urban living, reinforce community health programs, and allow interventions in managing air quality in time.

Ensemble learning for air quality index prediction: integrating gradient boosting, XGBoost, and stacking with SHAP-based interpretability

期刊论文
出版模式：

开放获取
发表日期：

2026年2月12日

查看全文

Sukhendra Singh,
Manoj Kumar,
Vishal Sengar,
Abhay Kumar,
Kumar Abhishek,
B. M. Ahamed Shafeeq

摘要

The increasing challenge of air pollution in cities requires smart methods to make proper predictions and manage the problem. Although machine learning and deep learning models have contributed greatly to weather and pollution forecasting, the main issue is the real-time flexibility, and scalability in the varying atmospheric conditions. This paper introduces a weighted Voting ensemble model that combines Gradient Boosting ( \(\times\) 4), CatBoost ( \(\times\) 3), XGBoost ( \(\times\) 2) and LightGBM ( \(\times\) 1) to improve the accuracy of Air Quality Index (AQI) forecasting. The full preprocessing (complete-case deletion, which retains extremes) and optimization of hyperparameters (GridSearchCV/Optuna, 5-fold CV) were used to enhance the robustness and generalizability of the model. The Taiwan Air Quality Dataset (2016–2024, \(n=4.6M\) hourly records from 74 stations), 6 major pollutants (PM2.5, PM10, \(\hbox {NO}_2\) , \(\hbox {O}_3\) , \(\hbox {SO}_2\) , CO), meteorological parameters (wind speed/direction), and 8-h averages) is used to model the data (spatial/temporal IDs are excluded, to allow deployment to a single station). Experimental validation of 60/16/24 splits (random + temporal validation) shows that the ensemble has validation MSE 0.6553 ( \(\hbox {R}^2\) 0.9969), which beats 15 baselines including the deep learning (LSTM MSE 45.4), but has temporal robustness ( \(\Delta\) \(\hbox {R}^2\) = − 0.0037). Moreover, SHAP is implemented to offer explainability, as it gives more insights into the contribution of features in predicting AQI. The results indicate the promise of interpretable ensemble learning systems to underpin sustainable urban living, reinforce community health programs, and allow interventions in managing air quality in time.