Hourly Particulate Matter (PM₁₀) Concentration Forecast in Germany Using Extreme Gradient Boosting

Abstract

Air pollution remains a significant issue, particularly in urban areas. This study explored the prediction of hourly point-based PM10 concentrations using the XGBoost algorithm to assimilate them into a geostatistical land use regression model for spatially and temporally high-resolution prediction maps. The model configuration and training incorporated meteorological data, station metadata, and time variables based on statistical values and expert knowledge. Hourly measurements from approximately 400 stations from 2009 to 2017 were used for training. The selected model performed with a mean absolute error (MAE) of 6.88 μg m−3, root mean squared error (RMSE) of 9.95 μg m−3, and an R² of 0.65, with variations depending on the siting type and surrounding area. The model achieved a high accuracy of 98.54% and a precision of 73.96% in predicting exceedances of the current EU-limit value for the daily mean of 50 μg m−3. Despite identified limitations, the model can effectively predict hourly values for assimilation into a geostatistical land use regression model.

Keywords

• air pollution
• machine learning
• XGBoost
• COSMO-REA6