Fuzzy Ensemble Algorithm for Day-ahead Photovoltaic Power Forecasting

Accurate photovoltaic power forecasting (PPF) is essential to optimize dispatch and balance generation and demand. This paper proposes a novel methodology called Fuzzy Ensemble Algorithm (FEA) for the day-ahead PPF based on Fuzzy C-means (FCM). FCM receives the day-ahead global horizontal irradiance (GHI) forecast as input, computes the membership matrix, and determines the centroids. Three specialist forecasting models (M2M3, and M4) based on long short-term memory (LSTM) are trained. Two complementary models (M1 and M5) are used to deal with extreme photovoltaic (PV) generation scenarios. The final forecast is obtained as a weighted sum of the predictions of two models, considering the probability vector. Furthermore, we propose an approach to transform the measured GHI into synthetic GHI forecast using a LSTM - autoencoder (AE), addressing the common challenge of the unavailability of day-ahead GHI forecasts in existing datasets. Our methodology is evaluated using a real-world dataset from a PV farm located at the State University of Campinas (UNICAMP) in Brazil. The empirical results show that the proposed methodology FEA outperforms Single model (SM) approach by more than 22% and 15% in terms of root mean squared error (RMSE) and mean absolute error (MAE), respectively.