Dispositivo de previsão de irradiância solar por meio de aprendizagem de máquina utilizando sensores de luminosidade LDR e imagens do céu

• Ano de defesa: 2020
• Autor(a) principal: Marinho, Felipe Pinto
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Energia solar; Aprendizagem de máquina; Processamento digital de imagens; Previsão de curto prazo
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/56028

One of the main difficulties in using the solar source to generate electricity for the grid is due to its intermittent behavior. Thus, short-term solar irradiance forecasts, which are those made for a horizons of until 6 hours ahead, are relevant so that the power grid operator can predict oscillations in the supply of solar plants and, in possession of this information, it can performed a better management of energy supply and demand. In this sense, in the present work, global solar irradiance forecasts were obtained for the forecast horizons of 10, 20 and 30 minutes a posteriori by the application of machine learning algorithms in data sets consisting of signals obtained by light dependent resistors and statistical descriptors (arithmetic mean, standard deviation and Shannon entropy) extracted from images of the sky captured by a camera, where the integration of such sensors was made using a Raspberry Pi 3. In this way, it was possible to evaluate whether the addition of predictors obtained from images of the sky provide an improvement in the performance of the applied machine learning models. In addition, it was also assessed if the use of the median smoothing and the Gaussian Laplacian sharpening filters over the images could increase the forecast accuracy relative to the case in which the statistical descriptors of the images are calculated without any filter. For the horizon of 30 minutes ahead, the arithmetic mean of the Root Mean Square Error obtained by the models which only the signals provided by the luminosity sensors were considered as attributes was 165.83 W / m², while for the case in which it is added information from unfiltered images it was obtained 154.76 W / m². For the case of adding information from images filtered by the median filter, a value of 154.08 W / m² was obtained and finally, for the other filter, a RMSE of 163.26 W / m² was obtained. Subsequently, the algorithms were applied to a new set of data that was built replacing the predictors related to the luminosity sensors by attributes characterized by being recurrence functions over the irradiance values of previous instants. For this case, for the 30 minute horizon, the mean of root mean square errors were 136, 43 W / m², 138.82 W / m², 135 W / m² and 140, 05 W / m² for cases without image, with image, image filtered with median and image filtered with the sharpening filter, respectively.