Machine learning framework for forecasting air pollution: Evaluating seasonal and climatic influences in Istanbul, Turkey

Abstract

Air pollution, driven by seasonal and meteorological variations, poses a significant threat to public health and urban sustainability. Despite numerous forecasting approaches, the influence of seasonal patterns on air pollutant levels remains underexplored. This study presents a computational framework utilizing the Nonlinear Autoregressive network with Exogenous inputs (NARX) model to predict concentrations of key pollutants SO₂, PM₁₀, NO, NOX, and O₃ in Esenyurt, one of the most industrialized districts in Istanbul, Turkey. Through systematic feature selection techniques, the study determines the most influential seasonal factors for each pollutant, reducing model complexity while improving predictive accuracy. The developed framework exhibits substantial improvements in predictive performance, with the optimal models achieving high determination coefficients (up to R²=0.965 for O₃) and low error metrics across training and validation datasets. Particularly, the inclusion of seasonal variables considerably improved prediction accuracy for NO, NO₂, and PM₁₀, while SO₂ predictions performed best when utilizing comprehensive seasonal indicators. These results demonstrate that seasonal dynamics play a crucial role in governing pollutant behavior and highlight the importance of incorporating such variables in forecasting models. This research contributes significantly to the field by advancing methodological approaches in air quality prediction while providing an adaptable model for policymakers and environmental agencies to implement in proactive pollution management strategies. Through examination of seasonal dependencies in air pollutant patterns, the study delivers a practical tool for urban planning and public health applications in rapidly expanding metropolitan regions.