Previsão de geração fotovoltaica utilizando redes neurais recorrentes

Abstract

In view of the growth in the use of photovoltaic (PV) solar energy in Brazil in recent years, it is important to be able to measure how much the plants will generate in the next few hours or days, so that the electrical sector can be planned reliably, guaranteeing the quality and continuous supply of electricity. Therefore, this work consisted of determining the best hyperparameters for building a PV solar energy forecast model using the Long Short-Term Memory (LSTM) recurrent neural network. After building the model, tests were carried out with the historical generation data of the solar plant in Pirapora - MG, with the objective of forecasting the generation values on a daily and hourly basis. Using the error metrics, Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE), values of 467.95 kWh, 329.25 kWh and 23.23% were found for the daily test and 41.21 kWh, 27.50 kWh and 36.07% in the hourly test, in which, for both cases, the test period was 6 months. In addition, RMSE, MAE and MAPE values were also obtained for the first 120 hours of the test, which were 19.47 kWh, 13.83 kWh and 16.05% for the period in question, respectively. The results found for the two forms of prediction were considered efficient, considering that only one input was used in the neural network.