A data fusion method to improve winter PM10 concentration predictions in Budapest based on the CAMS air quality models

Winter air pollution in Budapest is a major environmental issue, caused by an interaction of residential heating, urban traffic and large-scale transport. Increasing public and political demand are present to achieve more accurate air quality predictions to support both real-time public health measures and long-term mitigation policies.  Atmospheric chemistry and transport models of the Copernicus Atmospheric Monitoring Service (CAMS) provide near-real-time air quality forecasts for Europe. The validation of these model predictions for Budapest showed that although large-scale processes are well captured, the complex interaction of large-scale plumes with significant and highly variable local residential emissions leads to the underestimation of winter PM10 concentrations. Furthermore, CAMS models are not expected to fully predict the non-representative concentrations at specific urban monitoring locations, which, on the other hand, serve as the legal basis of all public policies and measures. Therefore, obtaining a relationship between monitoring site observations and CAMS model predictions is of primary importance. 

In this study, we used observed PM10 concentration data from 12 air quality monitoring sites within Budapest, as well as 24-hour predictions from 7 of the 9 CAMS models to produce an optimal linear combination of models that best matched, in terms of RMSE, the observed time series. A zero-degree term to correct the model bias was also applied. The applied data fusion method was cross-validated on urban monitoring sites not used in fitting the model, and found to improve PM10 forecast validation statistics compared to the pointwise model median (CAMS ensemble) as well as each of the 7 single models. The presented fusion of CAMS models can therefore provide an improved prediction of PM10 concentrations at urban monitoring sites in Budapest.