Otimização dos hiperparâmetros da LSTM por metaheurísticas para previsão de energia solar fotovoltáica

Abstract

THIS WORK INVESTIGATES THE APPLICATION OF METAHEURISTICS IN THE OPTIMIZATION OF LSTM NETWORK HYPERPARAMETERS FOR PHOTOVOLTAIC SOLAR ENERGY GENERATION FORECASTING. A DATABASE FROM A POWER PLANT IN NASIK (INDIA) WAS USED, COVERING THE PERIOD FROM MAY 15 TO JUNE 17, 2020. TWO METAHEURISTICS, GENETIC ALGORITHM (GA) AND FIREFLY ALGORITHM (FA), WERE EMPLOYED TO ADJUST PARAMETERS SUCH AS LEARNING RATE, REGULARIZATION, AND NETWORK STRUCTURE. THE MODELS WERE EVALUATED BY MSE, MAE, RMSE, AND COEFFICIENT OF DETERMINATION (R²), WITH GA ACHIEVING SUPERIOR PERFORMANCE COMPARED TO FA. THE RESULTS SHOW THAT OPTIMIZATION VIA AG PROVIDES MORE ACCURATE AND RELIABLE FORECASTS, HIGHLIGHTING THE IMPORTANCE OF THE APPROPRIATE CHOICE OF HYPERPARAMETERS AND THE QUALITY OF THE DATA SET TO IMPROVE FORECASTING MODELS.