Enhancing particulate matter prediction in Delhi: insights from statistical and machine learning models

This study advances our approach to modeling particulate matter levels-specifically, PM₁₀ and PM_2.5-in Delhi's dynamic urban environment through an extensive evaluation of traditional time series models (ARIMAX, SARIMAX) and machine learning models (RF, SVM) across air quality monitoring stations utilizing data from the period 2019 to 2023. We established a clear baseline of air quality variations using seasonal decomposition, highlighting critical seasonal peaks in PM₁₀ and PM_2.5 concentrations influenced by localized emissions and adverse weather conditions. Subsequent trend analysis revealed increasing PM₁₀ levels at several key monitoring stations, underscoring the impact of urban activities and seasonal variations. In contrast, reduction was observed in PM_2.5 levels at most monitoring stations. We utilized a wide range of exogenous variables, including other pollutants and meteorological parameters in our time series models to enhance the accuracy of predicting particulate matter. The SVM model proved to be more accurate in predicting particulate matter levels. It achieved testing RMSE values between 12.48 and 67.22 µg/m³ for PM₁₀ and 8.38 and 48.95 µg/m³ for PM_2.5, with testing R-squared values between 0.30 and 0.95 for PM₁₀ and 0.41 and 0.96 for PM_2.5. This research pioneers a methodologically enriched approach by systematically incorporating these exogenous factors, enhancing predictive capabilities, and deepening the understanding of complex environmental dynamics specific to urban cities like Delhi. The extensive spatial coverage and robust integration of diverse exogenous factors can significantly enhance environmental modeling, providing actionable insights for policymakers and advancing air quality forecasting in urban megacities.