Metodologia baseada em redes neurais artificiais para previsão da geração de plantas fotovoltáicas

Abstract

Precise predictions of photovoltaic generation (GPV) were investigated as significant reducers of the impac t s that intermittent sources of electrical generation provide to electrical systems. The GPV forecast involves uncertaint i e s that are related to the characteristics of the time series of solar resources and their highvolatility due to their dependence on many climatic conditions. Our paperpresents a comparative study between different applications of Artificial Neural Networks (ANN) applied to the hourly forecast of solar generation in two photovoltaic plants in the Alternative Energy Laboratory (LEA) of the Federal University of Ceara (UFC). In ourstudy, two configurations were used: a) Multilayer Perceptron (MLP); and b) Kohonen Neural Network (KNN), both trained with historical power data (W). Performance analysismetrics were applied to evaluate and compare the methodologies (correlation coefficient -r, Nash-Sutcliffe efficiency -NSE and relative bias -VR). The RNA results showed that applications with MLP had the best overall performance;SOMhad a slightly lower performance and also presented good potential for use in forecasting photovoltaic generation. Both ANNs applied had correlation close to 90%, low bias and NSE arou n d 0.80 and 0.90 in some applications.