Improved model output statistics of numerical weather prediction based irradiance forecasts for solar power applications

Abstract

The fast growth of solar photo-voltaic energy and the issues related to its integration in the power system are leading to an increased importance of forecasts of solar irradiance. Irradiance forecasts based on numerical weather prediction (NWP) models may be downscaled to finer spatial and temporal granularity and corrected for systematic biases by applying so-called model output statistics (MOS). This paper presents a MOS routine that is based on a large set of meteorological variables that are available from standard NWP output and a clear sky model. The method is based on a stepwise linear regression algorithm yielding a regression model with a set of variables that best explains the observed forecast error. The resulting irradiance forecasts for the first forecast day averaged over an ensemble of 27 stations corrected with this model reduces the relative root mean square error (rRMSE) to 22.7% compared to a rRMSE of 37.8% of uncorrected forecasts and a rRMSE of 25.6% of forecasts corrected with a method based on only the solar zenith angle and the predicted clear sky index – a method that is a current standard in NWP based irradiance forecasts. Furthermore, since this new method takes more meteorological information into account than the current standard method, the increase in skill evaluated in a probabilistic sense is even higher, because a forecast probability density is obtained that better reflects the sensitivity of forecast errors to atmospheric conditions.