Evaluation of statistical learning configurations for gridded solar irradiance forecasting

Gridded forecasts of solar irradiance are increasingly needed to integrate power into the electric grid from distributed solar installations and newer large-scale installations that don't have long records of observed irradiance. We evaluate different combinations of statistical learning models and aggregations of weather data from observed sites to identify which combination produces the lowest forecast errors at independent sites. The evaluation reveals how statistical learning model choice, closeness of fit to training data, training data aggregation, and interpolation method affect forecasts of clearness index at Oklahoma Mesonet sites not included in the training data. It shows that the choices of statistical learning model, interpolation scheme, and loss function have the biggest impacts on performance. Errors tend to be lower at testing sites with sunnier weather and those that are closer to training sites. All of the statistical learning methods and the NWP model output produce reliable predictions but underestimate the frequency of cloudiness compared to observations.