How good can we get? – An analysis of systematic errors in common models for all-sky imager based irradiance nowcasting for solar energy

Solar energy from photovoltaics (PV) is a major contributor to the power production, e.g., in Germany, with a growing share. It is a major contributor to renewable power production but highly volatile as it is heavily influenced by atmospheric conditions. Especially shading by clouds can change within seconds to minutes and cause ramps in irradiance and solar power production. Accurate short-term predictions (nowcasts) of irradiance for the next minutes can help to alleviate the impact of this volatility and improve the integration of solar power into energy grids. One approach for nowcasting is the use of all-sky imagers (ASI), ground based fisheye cameras which capture the current cloud situation. Therefore, cloud information is extracted from current images, future cloud states are extrapolated and converted into an irradiance nowcast. Despite substantial progress in the quality of the applied methods, current ASI nowcasting models still exhibit significant nowcast errors and struggle to reliably outperform persistence nowcasts for all situations. Therefore, we assessed the implications for nowcast performance of two common fundamental simplifications of ASI nowcasting models. Firstly, cloud evolution is often modelled by advection, i.e. simple displacement over time. Growth, shrinking or reshaping of clouds is usually neglected in the models. Additionally, the ASI viewing geometry may introduce a misrepresentation of the depicted cloud scene, which is also commonly neglected. The ASI views surrounding clouds from a single ground position and under varying angles. For direct irradiance however, the horizontal distribution of clouds and their intersection in the direction of the sun is essential. While ASI images are usually reprojected to comply with the required horizontal representation, the original difference in actual and required viewing geometry cannot be fully compensated. E.g., breaks between distant clouds may not be clearly visible by the ASI although modulating the irradiance. Kurtz et al. (2017) demonstrated a major impact by this geometric limitation. We applied a nowcasting model to synthetic ASI images of a simulated cloud scene to extend this previous study and analyze the errors introduced by both of the two commonly used simplifications of ASI nowcasting models. A large fraction of the nowcasting error is attributable to the simplifications, which implies a systematic baseline error of common ASI nowcasting models. While the implementation of more evolved cloud evolution and a better representation of relevant cloud geometry are challenging, this work indicates, that efforts to implement these improvements in ASI nowcasting models are a chance for a leap in performance of future nowcasting models.

 

Kurtz, B., Mejia, F., and Kleissl, J.: A virtual sky imager testbed for solar energy forecasting, Solar Energy, 158, 753–759, https://doi.org/10.1016/j.solener.2017.10.036, 2017.