Desenvolvimento de uma ferramenta para previsão de curto prazo da geração de energia fotovoltaica

Abstract

The dissemination of photovoltaic systems in the distribution of electricity, as a complement to the large centralized generating units, motivates research related to the forecast of the hourly capacity for distributed electricity generation. A good accuracy in the forecast of photovoltaic generation can contribute to the reduction of the variability resulting from the uncertainty of the output power of the photovoltaic system, improving its stability and dispatch. Thus, the proposal of new models that present a good predictive capacity is, therefore, of great interest to those responsible for the operation of photovoltaic generation systems. The present work aims to present a method of development of Artificial Neural Networks (ANNs) for the prediction of photovoltaic generation using the Multilayer Perceptron neural network model, in order to assist in the reliability of system operation, planning adjustment and allow for optimized dispatch. Local characteristics are the input information of the model that aims to analyze the influence of four meteorological variables: radiation (W / m²), ambient temperature (° C), wind speed (m / s) and humidity (%). All data used for training and testing the predictive ability of RNAs are real monitoring data, coming from a weather station and a photovoltaic generation plant. Generation forecasting capacity tests were carried out for a photovoltaic plant of 100 kWp for the period from January 2020 to June 2020, from the training of 60 RNAs in which the statistical evaluation of the results was carried out in relation to the variation of the number of neurons in a given range. The best ANN performance was defined from the analysis of the correlation coefficient and the metrics Mean Absolute Error (MAE) and Root Mean Square Error (RMSE). In order to validate the ANN's forecasting capacity, the assessment was carried out for each month separately and, finally, the sensitivity of each of the input variables in the performance of the proposed model was verified. In short, all the results obtained during the study are satisfactory and capable of proving the ability to predict the developed ANNs, this because, for all the cases under study, the correlation coefficient showed values greater than 0.9 which indicates a synaptic correlation very strong.