10.2.1. Short-Term Cloud-Motion Vector Forecasts
In standard resolution, short-term irradiance forecasts are produced using two consecutive satellite images, as discussed in Chapter 2 (see also Perez et al. 2002, 2004). Pixel-specific cloud motion is determined from these two images. The satellite images are first processed to remove solar-geometry effects. Each pixel is thus converted from sensor count to clear-sky index Kt*.2
The cloud-motion vector is then determined for each individual image pixel. The methodology used in SolarAnywhere is patterned after Lorenz et al. [7] (2007). Pixel-specific motion vectors are determined by calculating the RMSE of the difference between two consecutive image-derived Kt* maps surrounding the considered pixel when the second grid is advected in the direction of a motion vector. The selected motion vector corresponds to the lowest RMSE. This process is repeated for each image pixel, and each pixel is assigned an individual motion vector.
Kt* maps for subsequent hours, up to 6 h ahead,[8] are derived from localized motion. Future images are obtained by displacing the current image pixels in the direction of their motion vector. They are subsequently smoothed by averaging each pixel with its 8 surrounding intermediate-resolution neighbors representing an area of ~ 700 km2, following the pragmatic approach described by Lorenz et al. (2007).
The high-resolution (1 km) satellite images provide considerably more structural details than standard resolution (~ 10 km) satellite images. The cloud – motion vector approach can be applied to these images to generate subhourly scenes down to nearly 1 min.[9] Data streams of 1 min can be generated for (1) historical data, using motion-vector animation of two consecutive images and (2) future (forecast) data by projecting the latest image forward.
