Abstract
An estimated 28,000 premature mortalities are caused annually by adverse air quality in South
Africa. Accurate forecasts of PM2.5 concentrations equip people to avoid exposure at times of
particularly high concentrations. Despite the growing adoption of machine learning (ML) in air
quality forecasting globally, its adoption in South Africa remains limited.
This study investigates how well different ML regression models forecast PM2.5 concentrations
at lead times from 1-48 hours. The forecasts are made for four air quality monitoring stations
situated in the Limpopo and Mpumalanga provinces of South Africa near coal-fired power stations.
Meteorological and air pollution measurements from the monitoring stations near the
Medupi and Kusile power stations were used to train a Light Gradient Boosting Machine
(LGBM), a Random Forest, and an Extreme Gradient Boosting (XGBoost) model. These models
were evaluated using traditional ML metrics such as R2, Mean Absolute Error (MAE), and
Root Mean Squared Error (RMSE).
The results showed that short-term forecasting (1-3 hours ahead) achieved the best results
(R2 between 0.65 and 0.09), while performance declined gradually as the forecast horizon
increased. The LGBM model with Shapley Additive Explanations (SHAP) was the top performer
across various stations and time horizons, with the Marapong site downwind of Medupi
power station yielding the best results, and the Phola site in a low-income residential area
yielding the worst performance due to data quality, data availability, and complex emission
sources. The SHAP analysis revealed that engineered features, such as lagged PM2.5 concentrations
and rolling averages of PM2.5 concentrations, as well as meteorological variables
including temperature, air pressure, wind speed, and wind direction, were the most influential
factors in forecasting PM2.5 at various time horizons.
This study investigates the feasibility of using lightweight ML models to forecast short-term
PM2.5 concentrations in South Africa from historical measurements of meteorological variables,
PM2.5, and its precursors. Still, it depends significantly on the availability of good-quality
measurements. Given current data and machine learning models evaluated in this study,
ML models cannot yet match the performance of numerical models that simulate the underlying
physics and chemistry of the atmosphere. Future research should focus on enhancing
data availability, integrating multiple data sources such as satellite imagery, and exploring
deep learning models to improve medium to long-term forecasting.