Previsão de geração fotovoltaica a partir de dados meteorológicos utilizando rede LSTM

Abstract

The stochastic and nonlinear nature of some meteorological factors, such as the local solar radiance, ambient temperature, wind speed, directly influence photovoltaic energy generation. Furthermore, estimating the produced energy amount is fundamentally important to production quality, which can avoid problems on the electrical network. Based on this problem, this article presents an application of LSTM (LongShort-Term Memory) networks for forecasting generation (power) data on photovoltaic panels using meteorological data. The data were collected from a low-cost solarimetric station close to the production plant. The obtained information includes solar radiance, panels temperature, ambient temperature, humidity, wind speed, rain amount, voltage, and current. It is used to feed a neural network, whose function is the prediction of the next five seconds of power produced by the solar panels. Results show that the LSTW Network presents good estimation values, having a low mean absolute error for each second.