When considering exogenous variables for the forecasting of solar irradiance and related phenomena, the most important one is sky condition and, in particular, cloud cover. Cloud-cover information can be obtained from satellite images (e. g., http://www. goes. noaa. gov/browsw. html) or from ground – based sky imagers—usually a CCD camera taking pictures of a convex mirror that reflects the entire sky, or one fitted with a fisheye lens pointed upward. Remote sensing provides images with very large fields of view (worldwide if necessary) but with medium spatial and temporal resolutions, whereas sky imaging provides images of high spatial and temporal resolution for small fields of view (usually not more than 10-20 km). Field of view is a very important parameter given that it determines the maximum forecasting horizon for which the images are useful. Typical ground-to-sky-imaging techniques provide no information for time horizons greater than 30 min (see Chapter 9).
In this section, we study the effect of adding exogenous variables derived from sky-image processing. We do not use advanced machine-learning techniques such as ANN because our main goal is to highlight the usefulness of exogenous variables in the forecasting of solar irradiance. In this case, the objective is to use information derived from sky images to generate short-term forecasts of DNI at ground level. Specifically, we are interested in forecasting 1 min averaged DNI values for time horizons varying 3-15 min. The solar forecasts derived here are analyzed and quantified in terms of RMSE deviations in relation to actual values, and compared to the performance of the dull- persistence model.