Estimativa de irradiância solar utilizando redes neurais artificiais e imagens do céu

Abstract

The growing global demand for electricity from renewable sources is increasingly encouraging the adoption of solar plants. However, due to the variant nature of solar irradiance, brought about by atmospheric phenomena, solar plants are not able to generate energy continuously. The short-term estimation of solar irradiance becomes a fundamental task for the proper planning of such plants, as well as systems connected to domestic networks. Commercially, devices capable of periodically performing readings of solar irradiance are available, however, the adoption of such devices becomes prohibitive due to the high cost for acquisition and maintenance. In this scenario, studies present different approaches to estimate solar irradiance, with emphasis on the use of artificial neural networks. This work subjectively reviews some of these studies and proposes to present a model of artificial neural networks to estimate solar irradiance from images of the sky. The proposed methodology begins with data collection and preparation, establishing which information would contribute more significantly to the network learning. Following, a total of 10 configurations were designed and tested, containing a varied number of layers and neurons. The accuracy of these settings is verified from the mean quadratic error and direct comparison with the values measured by a luximeter. The network configuration that presented more accurate results was the model with 30 neurons in 1 inner layer. The data set and the number of entries established for the neural network were satisfactory.