Machine learning approaches for short- and multi-day PM2.5 prediction in Tabriz, Iran

Accurate forecasting of fine particulate matter (PM 2.5 ) is crucial for effective urban air quality management and public health protection. This study develops machine learning models to predict PM 2.5 concentrations at three urban monitoring stations in Tabriz, Iran, using hourly data from 2013 to 2023 and external validation with 2024 measurement results. Both short-term (6, 12, 24 h ahead) and multi-day (24, 48, 72 h averaged) predictions were performed. Lagged PM 2.5 values, short-term temperature forecasts, and daily natural gas consumption were incorporated to capture temporal and anthropogenic effects, while data from neighboring stations accounted for spatial dependencies. Tree-based boosting algorithms, particularly Extreme Gradient Boosting (XGBoost) and Categorical Boosting (CatBoost), outperformed other models for short-term forecasts (e.g., R 2 = 0.82, RMSE = 6.59 μg/m 3 for 6-h prediction), whereas simpler models, including Linear Regression and Neural Networks, achieved comparable RMSE for longer horizon. Feature importance analysis identified ambient temperature, recent PM 2.5 concentrations, and gas consumption as dominant predictors. The models maintained robust performance under external validation, demonstrating their potential for operational air quality forecasting and urban pollution mitigation.