Explainable artificial intelligence-driven model for ultrafine particle (PM0.1) prediction and explanation using meteorological variables

Abstract

PM_0.1, an ultrafine urban air pollutant, poses significant health risks due to its ability to penetrate deep into the lungs, enter the bloodstream, and rapidly circulate throughout the human body, potentially causing severe respiratory diseases. Effective monitoring and explanation of PM_0.1 concentrations are essential for implementing targeted air quality management strategies. This research addresses these challenges through a novel soft computing framework that integrates multiple computational intelligence techniques for both accurate prediction and interpretation of PM_0.1 concentrations using meteorological variables. Our methodology employs boosting algorithms with hyperparameter optimization to develop the most effective predictive model. Validated using real-world urban air quality data from Southeast Asia, our framework demonstrates broad applicability for diverse urban environments. Performance evaluation using unseen data demonstrates that XGBoost with Bayesian optimization achieves superior results, with an R² of 89 %, Mean absolute error (MAE) of 0.28 µg/m³, and Root Mean Squared Error (RMSE) of 0.38 µg/m³. By integrating SHAP (SHapley Additive exPlanations) with the XGBoost model, we provide global and individual explanations of the model's predictions, highlighting the relative importance of meteorological variables such as temperature, humidity, and wind speed. This hybrid soft computing approach enhances the model's practical utility for environmental decision-making. It sets a foundation for future intelligent environmental monitoring applications across diverse regions, offering a scalable solution for targeted urban air quality management that balances prediction accuracy with interpretability.