Hybrid FFT-Hyperband-LSTM model for direct short-term PV power forecasting

Abstract

This paper deals with the direct short-term prediction of photovoltaic power using a hybrid model consisting of the Hyperband-FFT-LSTM hybrid model. The Hyperband model is used to optimise the hyperparameters of the LSTM model, the Fast Fourier Transform (FFT) is used to extract the most important frequencies from the data and the LSTM model to perform the actual prediction. The model was trained in Python 3 with Tensorflow using data such as photovoltaic power, irradiation, temperature and wind speed generated by the PVGIS simulator. In terms of forecasting performance, a root mean square error (RMSE) of 1.5 MW, a mean absolute error (MAE) of 1.23 MW and a coefficient of determination (R2) of 0.74 were obtained for a 72-hour forecast. In order to validate the model, it was tested on an existing photovoltaic power plant with a peak power of 25 MWp and its forecasts were compared with the power recorded on a power plant. Overall, the model showed good accuracy.