Estimation and prediction of cloudiness from ground-based all-sky hemispherical digital images

Cloudiness is the environmental factor most a ecting the solar radiation reaching a particular location on the Earth surface. Therefore it has great impact on the performance of predictive solar radiation models for that location. This work aims contributing to the improvement of such models by using ground-to-sky hemispherical colour digital images as a means to estimate the fraction of visible sky corresponding to clouds and to clear sky, and by using radial basis function neural networks to model and predict the estimated cloudiness time-series. The general approach for cloudiness estimation, common to many image processing applications, consists in nding one threshold on a given pixel intensity scale that segments the image pixels into clear sky and cloud. The neural network models, selected by means of multiobjective genetic algorithms, are trained as one-step-ahead predictors and used iteratively in order to predict the cloudiness time-series up to a required prediction horizon. In order to allow the evaluation and comparison of image thresholding methods as well as forming a timeseries suitable to train and evaluate neural network models, the pixels of 410 images were manually classified as clear sky or cloud, establishing a reference database. Two well known image thresholding algorithms are tested and a neural network approach is presented. For the latter, a number of statistical measures are extracted from the images constituting a feature space of potential inputs. As well as for the cloudiness predictive models, the actual inputs and number of neurons to be employed are selected by means of a multi-objective genetic algorithm