Multivariate Analysis for Solar Resource Assessment Using Unsupervised Learning on Images from the GOES-13 Satellite

Solar resource assessment is of paramount importance in the planning of solar energy applications. Solar resources are abundant and characterization is essential for the optimal design of a system. Solar energy is estimated, indirectly, by the processing of satellite images. Several analyses with satellite images have considered a single variable—cloudiness. Other variables, such as albedo, have been recognized as critical for estimating solar irradiance. In this work, a multivariate analysis was carried out, taking into account four variables: cloudy sky index, albedo, linke turbidity factor (TL2), and altitude in satellite image channels. To reduce the dimensionality of the database (satellite images), a principal component analysis (PCA) was done. To determine regions with a degree of homogeneity of solar irradiance, a cluster analysis with unsupervised learning was performed, and two clustering techniques were compared: k-means and Gaussian mixture models (GMMs). With respect to k-means, the GMM method obtained a smaller number of regions with a similar degree of homogeneity. The multivariate analysis was performed in Mexico, a country with an extended territory with multiple geographical conditions and great climatic complexity. The optimal number of regions was 17. These regions were compared for annual average values of daily irradiation data from ground stations using multiple linear regression. A comparison between the mean of each region and the ground station measurement showed a linear relationship with a R2 score of 0.87. The multiple linear regression showed that the regions were strongly related to solar irradiance. The optimal sites found are shown on a map of Mexico