This paper presents the development and evaluation of a probabilistic uncertainty model for the CAMS radiation service irradiance estimates. This model is based on the conditioning of the cumulative distribution function of the deviations to the model inputs. It provides for each value of the historical time series and at each selected geographical location an independent probabilistic uncertainty.
The Copernicus Atmosphere Monitoring Service (CAMS) provides historical time series of global horizontal, beam normal, and diffuse horizontal irradiance in Europe, Africa, South America, and the Asia-Oceania region trough the CAMS Radiation Service (CRS). The CRS was widely validated and is routinely monitored on the basis of a large number of ground observation stations. Nevertheless, existing studies all derive average validation metrics, which can only be seen as typical, aggregated results. Users get no information about the uncertainty of the estimate at individual data points. This study systematically scans through a large database of deviations between the CRS irradiance estimates and ground-based irradiance observations. A look-up-table that describes the uncertainty of the CRS deviations is obtained by conditioning the cumulative distribution function of the deviations to the CRS model inputs. Parametric and nonparametric probabilistic representations of this uncertainty model are investigated. This model provides a probabilistic deviation for each value of the CRS estimate time series and at each geographical location. The uncertainty model shows very good calibration and sharpness metrics in all sky conditions as well as an average Continuous Ranked Probability Score of 50 W/m2$^2$. Additionally, its very strong robustness against the selection of training stations is shown.
Solar RRL, EarlyView. Read More
