Time Series Forecasting of PM 2.5 Concentration and Associated Health Risks Using a Supervised Hybrid Machine Learning (SHML) Model

The persistent release of PM 2.5 emissions from various sources has degraded air quality, causing environmental and human health issues. Managing elevated PM 2.5 levels is challenging due to limited monitoring, inadequate emission inventory, and vulnerability to multicollinearity and nonlinearity, underscoring the need for robust and reliable models. This study developed a Supervised Hybrid Machine Learning (SHML) model combining Support Vector Regression (SVR) and Artificial Neural Network (ANN) to predict daily and monthly PM 2.5 concentrations. The SHML model was developed on four years of hourly PM 2.5 and meteorological data, and its performance was evaluated against baseline models by estimating R² and mean absolute error (MAE). Shapley Additive exPlanation (SHAP) assessed meteorological factors' impact on model accuracy. Human health risks and environmental cancer burden from PM 2.5 inhalation were also estimated. The SHML model performed better, with R² = 0.921 and MAE = 0.002. Meteorological factors like precipitation, wind speed, and temperature significantly influenced the model. Non- carcinogenic health risks were 7.02 to 8.33 times higher than the acceptable threshold (<1), and carcinogenic risks exceeded acceptable levels (1×10–4 to 1×10–6). Similarly, lung cancer mortality attributable fraction (35-38%) was higher than cardiopulmonary mortality (25- 27%). The study recommends the SHML model for PM 2.5 prediction because it captures temporal dependencies, multicollinearity, nonlinearity, and superior results.