An Evaluation of Statistical Learning Methods for Gridded Solar Irradiance Forecasting

As solar power usage continues to grow, the need for accurate solar irradiance forecasts over large areas also increases. Most operational solar forecasting systems combine numerical weather prediction model output with statistical learning models to produce site-based forecasts calibrated to an archived observation dataset. The weakness of this approach is that accurate forecasts cannot be made at a new site until a new archive of observations is generated. A gridded solar irradiance forecasting system offers the advantage of aggregating the information from multiple sites to produce calibrated forecasts on a uniform grid. Within the framework of a gridded forecasting system, there are many different choices that can be made in the process of selecting input variables, statistical learning models, and grid application methods. This study uses the National Center for Atmospheric Research (NCAR) Gridded Atmospheric Forecast System (GRAFS) to test different solar irradiance forecasting approaches within a common framework. One approach is to train a single statistical learning model using data aggregated from multiple sites and then apply that model at every grid point in the domain. Another approach is to perform statistical learning corrections at each site and then use a spatial interpolation model to estimate the values at each grid point. This study compares both approaches over a multi-month period with observations from the NOAA MADIS system and the Oklahoma Mesonet. Verification statistics and case studies are shown.