Predição da geração de energia fotovoltaica aplicando o modelo Narx

Abstract

The intermittent nature of sunlight creates an imbalance between the generation of photovoltaic energy and the consumption of electricity, making it difficult to control the entire electrical system. Therefore, for better controlling it’s generation, a highly accurate estimate in the next hour or on the business day of power generation, for allowing better management and commercialization of photovoltaic electricity. For this reason the nonlinear autoregressive network with exogenous inputs NARX was applied in this research, aiming to select the an effective network configuration for predicting photovoltaic energy generation. Data from the meteorological station, such as temperature and solar radiation, were used in addition to the energy generated from a private photovoltaic microgeneration in Varginha - MG, during the period between 08/04/2019 and 20/04/2020. Data from two other micro-generation plant were compared for network performance analysis. The three databases were used as input variables in the NARX network, where different numbers of delays and neurons were applied, totaling thirty configurations. Through the mean squared errors of each configuration, analysis of variance and multiple comparison Scott Knott test were performed to select the most efficient network configuration for all three databases. It was concluded that the NARX network is suitable for predicting the generation of photovoltaic energy hour ahead. The results obtained based on the mean quadratic error (EMQ = 0.0027) and determination coefficient (R2 = 0.979) showed that the configuration of 50 delays and 15 neurons showed greater efficiency in the prediction. Polynomial regression was applied to database 1, obtaining the coefficient of determination (R2 = 0.9166).