The need for accurate solar irradiance and power forecasting is growing as more solar power enters the grid. There are two basic forecasting approaches used: (1) numerical weather prediction (NWP) that forecasts cloud coverage and irradiance through the dynamical modeling of the atmosphere and (2) models that use current conditions and recent trends in clouds to infer the impact on irradiance in the near future. The NWP approach tends to produce more accurate forecasts for look-ahead periods of 6 hours and longer. The current conditions and recent trend approach tends to produce better forecasts for the first 3-5 hours ahead depending on location. A recent comparison by Perez et al. 2011 looked at the day-ahead forecast performance of various global scale and mesoscale NWP models. In this study it was found that the global scale models provided forecasts with lower RMSE than the mesoscale models run at higher resolutions as shown in Figure 1.

Fig. 1. Composite RMSE of Day 1 and Day 2 irradiance forecasts for all SURFRAD sites forecasted for the period May 1, 2009 – April 30, 2010 (Perez et al. 2011).

Presumably running higher resolution mesoscale models should produce more accurate day-ahead forecasts than the coarse resolution global scale models.  Since improving the day-ahead forecast is important to grid managers and since running mesoscale models takes significant computational resources, it is important to determine why the global scale model outperformed the mesoscale models in the Perez et al. 2011 study. The current research and the presentation will focus on determining the reason for the lower RMSE for the global models as compared to the mesoscale models. Several possibilities for the accuracy differences will be investigated, including:

             (1) the accuracy of the initial and boundary conditions ingested by the mesoscale models (the initial and boundary conditions for all mesoscale models were from the same model),

            (2) the possibility that the mesoscale model radiation schemes are not sophisticated enough to capture the detail needed for accurate irradiance forecasts, and

            (3) the statistical postprocessing performed on the global scale models improved the accuracy of the global scale forecasts.