Ensemble learning for air quality index prediction: integrating gradient boosting, XGBoost, and stacking with...

Area of Science:

• Environmental Science
• Data Science
• Computer Science

Background:

• Urban air pollution poses a significant challenge, necessitating advanced prediction and management strategies.
• Existing machine learning and deep learning models struggle with real-time flexibility and scalability in dynamic atmospheric conditions.

Purpose of the Study:

• To develop a robust and accurate Air Quality Index (AQI) forecasting model.
• To enhance the real-time flexibility and scalability of air pollution prediction systems.

Main Methods:

• A weighted Voting ensemble model combining Gradient Boosting, CatBoost, XGBoost, and LightGBM.
• Comprehensive data preprocessing and hyperparameter optimization using GridSearchCV/Optuna with 5-fold cross-validation.
• Utilized the Taiwan Air Quality Dataset (2016-2024) encompassing pollutants, meteorological data, and hourly records from 74 stations.

Main Results:

• The ensemble model achieved a validation Mean Squared Error (MSE) of 0.6553, significantly outperforming 15 baseline models, including LSTM (MSE 45.4).
• Demonstrated temporal robustness with a Δ R² of -0.0037.
• SHAP analysis provided feature importance insights for enhanced model interpretability.

Conclusions:

• The proposed interpretable ensemble learning system shows significant promise for improving urban air quality management.
• The findings support sustainable urban living, community health programs, and timely air quality interventions.