This paper examines the skill of three different WRF set-ups (WRF-dudhia, WRF-rrtmg, WRF-solar) compared to it’s driving data, GFS. The aim is to produce an hourly day-ahead forecast for Singapore. A multivariate post-processing technique which combines principal component analysis (PCA) with stepwise variable selection is applied to all four models. These are also compared to smart persistence and a climatological forecast. WRF-solar combined with PCA and stepwise selection method produced the best results.
In this study the output is average for the whole island of Singapore. Clear sky index is used to estimate the forecast as it removes the diurnal and seasonality in irradiance. The three steps to the post-processing method are: 1) removal of the yearly and daily cycles, 2) dimensionality reduction and 3) Model Output Statistics (MOS). Stepwise variable selection is used to select the best possible set of explanatory variables. PCA is used for dimensionality reduction of the 3D model output. 80% of the explained variance is accounted for in at most 5 PCs. This method is compared to 1) no dimensionality reduction and 2) 3D variables are averaged vertically.
All post-processing techniques improve the forecast with WRF-solar-PCA performing best. Although there are certain months when other configurations perform well. Different models perform better during different weather conditions (Fig. 1)