Improving air quality prediction using bayesian optimization for feature selection, hyperparameter tuning and ensemble learning

Air pollution from fine particulate matter (PM2.5) is a continuing environmental and public health issue in Malaysia, where seasonal haze and local emissions cause air quality to fluctuate considerably. Most machine learning models for predicting PM2.5 rely on the feature selection and hyperparameter tuning stage as two separate steps, instead of optimising two simultaneously, limiting the accuracy and generalization of PM2.5 forecasting. This study develops an enhanced prediction model applying Bayesian Optimization (BO) to feature selection and hyperparameter tuning simultaneously, within a stacking ensemble framework. Daily air quality records for Kuching, Sarawak (3,026 entries, 2016-2024) were pre-processed into 2,557 cleaned observations, The input features used to predict PM2.5 were PM10, nitrogen dioxide (NO₂), carbon monoxide (CO) and ozone (O₃). BO was used to determine the most influential feature subset and optimize four base learners: Random Forest, XGBoost, CatBoost and Support Vector Regression. Evaluation used RMSE, MAE and R² under a chronological train-test split. All the base learning models performed better after optimization, but SVR was most successful with the highest enhancement. Its coefficient of determination (R²) increased from 0.7790 to 0.9051, while the Mean Absolute Error (MAE) improved from 5.03 to 3.56. The stacking ensemble had a test R² of 0.9276, RMSE of 2.4534, and MAE of 1.5983. The robustness of the stack was also tested for five different train-test split ratios and the mean R² was 0.9164 ± 0.0273. The predicted pollutant concentrations were also converted into AQI categories using the EPA breakpoint level and the classification of the AQI was 90.04% for 512 test samples. The final forecasting model was also integrated into a prototype web-based dashboard to offer a practical tool to help with decision making. The main contribution of this study is a systematic and reproducible framework in which feature selection and hyperparameter tuning are optimised jointly through Bayesian Optimization, helping environmental authorities and the public to predict and respond to poor air quality.