Improving the accuracy of hourly satellite-derived solar irradiance by combining with dynamically downscaled estimates using generalised additive models. (October 2016)

• Record Type: Journal Article
• Authors: Davy, Robert J.; Huang, Jing R.; Troccoli, Alberto
• Is Part Of: Solar energy. Volume 135(2016)
• Page Start: 854
• Page End: 863
• Publication Date: 2016-10
• Subjects: Clear sky index -- DNI -- GHI -- Insolation; Solar energy -- Periodicals; Solar engines -- Periodicals; 621.47
• DOI: 10.1016/j.solener.2016.06.052
• Languages: English
• ISSNs: 0038-092X

Abstract

Highlights: The accuracy of hourly satellite irradiance is enhanced using statistical models. Generalised additive models with smoothing splines outperform polynomial regression. Performance improves further when weather model irradiance is used as a predictor. Spatial weighting of regression models was found to increase the risk of bias. Abstract: The gridded hourly solar irradiance derived from satellite imagery by the Australian Bureau of Meteorology represents the current state of the art in quantification of the long-term solar resource for locations where no ground measurements are available. Using nonparametric regression, we test the potential for the satellite-derived global horizontal irradiance and direct normal irradiance to be improved by combining with irradiance that has been dynamically downscaled using a numerical weather prediction (NWP) model. NWP irradiance, together with the satellite irradiance, solar zenith angle and their interaction terms are used as inputs to generalised additive models (GAM) using smoothing splines. The spatial weighting of these empirical models according to distance is also tested. Cross validation with ground measurements indicates that RMSE can be improved by a few percent over the satellite-derived irradiance. The addition of dynamically downscaled irradiance as a GAM predictor further improves RMSE by a few percent, depending on location. When these empirical models are weighted spatially over large distances (hundreds of kilometres), the results are more equivocal. However, spatial weighting of the regression functions should be possible over smaller regions where the atmospheric turbidity properties are similar.