Multi-metric evaluation of machine learning models for PM2.5 forecasting in diverse urban regions of Gujarat, India

Severe PM 2.5 pollution in India highlights the need for reliable forecasting and timely interventions due to its serious health and environmental impacts. This study evaluates five machine learning models — Support Vector Regression, Gaussian Process Regression, Random Forest, XGBoost, and Stacking — for next-day PM 2.5 forecasting across three contrasting urban environments in Gujarat: Sector 10 (green zone) in Gandhinagar, and Maninagar (densely populated) and Vatva (industrial zone) in Ahmedabad. All models demonstrated comparable predictive performance, with Random Forest achieving slightly better accuracy (RMSE: 12.30; MAE: 8.62). However, Random Forest and XGBoost exhibited overfitting tendencies despite extensive hyperparameter tuning, while Stacking provided no substantial improvement because of the high correlation among base-model predictions. Sector 10 showed stronger forecasting performance than Maninagar and Vatva, while threshold-detection performance was limited because PM 2.5 exceedances were relatively rare. SHapley Additive exPlanations analysis identified current-day PM 2.5 as the dominant predictor, while an additional feature exclusion experiment demonstrated the substantial contribution of autoregressive information. An additional Random Forest analysis excluding current-day PM 2.5 , its lagged terms, and PM 10 substantially reduced predictive performance across all three regions, with the overall average RMSE increasing from 12.30 to 18.39. Taylor diagrams further showed that the models captured general temporal patterns but underestimated observed variability. These findings highlight both the predictive capability and limitations of machine learning approaches for short-term PM 2.5 forecasting. Future research should consider broader spatial coverage, additional atmospheric and emission-related predictors, classification-based exceedance prediction, stronger regularisation, and advanced deep learning approaches.