Predictive Models for Urban Air Quality Management Using AI

Abstract

Air pollution threatens public health in datascarce urban areas like Sri Lanka, where sparse monitoring hinders proactive management. We propose an integrated AI framework: hybrid SARIMAX-Temporal Fusion Transformer for multi-pollutant forecasting, ensemble spatial estimation for gap-filling, CEEMDAN-Seq2Seq for 24-hour AQI risk alerting, GRU for anomaly detection, and XAI for transparency. Validated on Central Environmental Authority data (20192024), the model achieves an 81.6% decrease in the value of the RMSE metric for ozone forecasting, as well as an R2 value of 0.9077 for high-risk AQI prediction, outperforming the baseline methods by 15-81%. The framework is modular in nature, thereby providing policymakers with the ability to use real-time dashboards, thus making Sri Lanka move from reactive to proactive management.