Spatiotemporal bias correction of CAMS PM₂.₅ forecasts over Europe via ensemble machine learning

Accurate spatiotemporal forecasts of fine particulate matter (PM₂.₅) remain a central challenge in air quality modelling, particularly for forecast horizons exceeding 24 hours. This study presents a hybrid, data-driven framework designed to correct systematic biases in regional forecasts from the Copernicus Atmosphere Monitoring Service (CAMS) across Europe.The modelling framework integrates multiple data sources describing atmospheric composition, meteorology, emissions, and spatial context. Predictor variables include hourly CAMS forecasts for key pollutants (PM₂.₅, PM₁₀, O₃, NO₂), ERA5-Land meteorological fields, boundary layer height (PBLH) from ERA5, EMEP emission inventories, and static predictors such as population density, elevation, and land-use fractions. Ground-based PM₂.₅ observations from 1510 monitoring stations across EEA member countries serve as the target variable. Predictive performance is evaluated for two forecast horizons, +24 and +48 hours, using three machine learning algorithms: Extremely Randomized Trees (ET), Extreme Gradient Boosting (XGBoost), and Multi-Layer Perceptrons (MLP). Models are trained independently and validated using a spatiotemporal cross-validation scheme and subsequently combined into an ensemble model. Training is conducted on data from 2023–2025, with hyperparameters optimized via Bayesian optimization on a stratified 10% subset of monitoring stations.Results show that all individual models consistently outperform the raw CAMS forecasts across multiple metrics, including R², RMSE, and bias, for both +24 and +48 hours’ time leads. The ensemble further improves predictive performance relative to both CAMS and individual learners. Performance gains are observed across regions and seasons, indicating improved representation of spatial heterogeneity and short-term pollution dynamics. The framework effectively reduces systematic CAMS biases, particularly in regions characterized by complex emission structures and heterogeneous land use. Variable importance analysis highlights CAMS predictors, PBLH, and residential emissions as key drivers of model performance.The proposed framework demonstrates the potential of combining physically based forecasts with machine learning-based bias correction to improve operational air quality predictions at the continental scale. By leveraging heterogeneous data sources and a rigorous validation design, the approach provides a scalable and interpretable solution for enhancing PM₂.₅ forecasts across Europe.