Classifying direct normal irradiance 1-minute temporal variability from spatial characteristics of geostationary satellite-based cloud observations

Abstract

Variability of solar surface irradiances in the 1-minute range is of interest especially for solar energy applications. Eight variability classes were previously defined for the 1 min resolved direct normal irradiance (DNI) variability inside an hour. In this study spatial structural parameters derived fromsatellite-based cloud observations are used as classifiers in order to detect the associated direct normal irradiance (DNI) variability class in a supervised classification scheme. A neighbourhood of 3×3 to 29×29 satellite pixels is evaluated to derive classifiers describing the actual cloud field better than just using a single satellite pixel at the location of the irradiance observation. These classifiers include cloud fraction in a window around the location of interest, number of cloud/cloud free changes in a binary cloud mask in this window, number of clouds, and a fractal box dimension of the cloud mask within the window. Furthermore, cloud physical parameters as cloud phase, cloud optical depth, and cloud top temperature are used as pixel-wise classifiers. A classification scheme is set up to search for the DNI variability class with a best agreement between these classifiers and the pre-existing knowledge on the characteristics of the cloud field within each variability class from the reference data base. Up to 55 % of all DNI variability class members are identified in the same class as in the reference data base. And up to 92 % cases are identified correctly if the neighbouring class is counted as success as well - the latter is a common approach in classifying natural structures showing no clear distinction between classes as in our case of temporal variability. Such a DNI variability classification method allows comparisons of different project sites in a statistical and automatic manner e.g. to quantify short-term variability impacts on solar power production. This approach is based on satellite-based cloud observations only and does not require any ground observations of the location of interest.