A Transformer-Based Framework for Accurate and Interpretable PM2.5 Forecasting through Integration of Country-Level Embeddings and Explainable AI for Enhanced Environmental Decision-Making

Air quality prediction is increasingly vital to environmental health, particularly for urban and rural regions, where dangerous particulate matter hazards often exist. This study presents a novel methodology for estimating PM2.5 concentrations using a transformer model with country-level embedding and explainable AI (XAI). The proposed approach is superior to conventional machine learning and deep learning techniques as it provides high accuracy and is interpretable and applicable to various geographic regions. Given the country-specific embeddings, the transformer architecture models the replacements caused by time and location variations in pollutants' concentrations, consequently allowing accurate predictions even for regions having sparse data. Furthermore, SHAP and LIME elucidate the model's tendency to predict, providing policymakers with valuable insights. Overall, the proposed architecture presents a stronger predictive power than other forecasting models, with an R-squared value of 0.98 and a mean absolute error of 0.011. Also, using country embedding has helped improve accuracy and the ability to apply to different regions. Hence, this research offers a plausible framework to forecast air pollution and evidence-based government policymaking and planning about air pollution and its health and environmental effects.