One-hour-ahead solar radiation forecasting by MLP, LSTM, and ANFIS approaches

Creative Commons License

Yıldırım A., Bilgili M., Özbek A.

METEOROLOGY AND ATMOSPHERIC PHYSICS, vol.135, no.1, pp.1-17, 2023 (SCI-Expanded)


The use and importance of renewable energy sources (RES) have been increasing every passing year as fossil fuels will
soon be depleted. Within this context, solar-photovoltaic (PV) is the most preferred energy type among RES. The PV has
uncertain power output as its output depends on solar radiation, which is heavily influenced by environmental factors, so
the prediction of solar radiation plays a crucial role in integrating these plants into the electricity grid. For the short-term
1-h-ahead solar radiation prediction, four time-series methods were implemented in this study: long short-term memory
(LSTM) network, multilayer perceptron (MLP), and adaptive neuro-fuzzy inference system (ANFIS) with grid partition
(GP), and fuzzy c-means (FCM). Root mean square error (RMSE), correlation coefficient (R), and mean absolute error
(MAE) were used as statistical error criteria. The obtained results by the LSTM, MLP, ANFIS-FCM and ANFIS-GP models
were assessed by comparing with the actual data. Considering the testing procedure, the best MAE values were found to be
53.37 W/m2, 58.45 W/m2, 61.68 W/m2, and 78.17 W/m2 for the LSTM, ANFIS-FCM, MLP, and ANFIS-GP, respectively.
Results showed that the LSTM model in 1-h-ahead solar radiation prediction yielded the best results among all four models
with high accuracy.