Solar power prediction based on Artificial Neural Network guided by feature selection for Large-scale Solar Photovoltaic Plant

• Published in: Energy Reports
• Main Author: Amer H.N.; Dahlan N.Y.; Azmi A.M.; Latip M.F.A.; Onn M.S.; Tumian A.
• Format: Article
• Language: English
• Published: Elsevier Ltd 2023
• Online Access: https://www.scopus.com/inward/record.uri?eid=2-s2.0-85173601977&doi=10.1016%2fj.egyr.2023.09.141&partnerID=40&md5=3a3deab2d430dad7f7668a572ca29240

Malaysia's most plentiful renewable energy (RE) source, solar, only generate power during the day. High penetration of intermittent PV generation, especially from Large-Scale Solar Photovoltaic (LSSPV) plants, may complicate power system planning and operation. Hence, the need to accurately forecast solar generation for sustaining grid stability. This study identified the meteorological variables that significantly impact the power generation of an LSSPV plant using the Pearson Correlation Coefficient (PCC). The LSSPV plant, located in the central area of Malaysia, is used as a case study. To this end, a regression analysis model was developed using Artificial Neural Network (ANN) to predict solar generation with changing meteorological conditions. From August 1, 2020, to July 1, 202,1 and September 1, 2022, to November 1, 2022, a 5-minute dataset is collected. The meteorological variables with the strongest correlation on solar generation were the total global horizontal irradiance and global irradiance on the module plane selected by PCC, followed by PV module temperature, ambient temperature, wind speed, total slope and total horizontal irradiation. The final input to the forecasting model consists of four primary features that indicate a significant correlation with solar PV generation, which are total global horizontal irradiance (W/m²), PV module temperature (°C), ambient temperature (°C) and wind speed (m/s). The simulation findings reveal that the accuracy evaluation metrics are higher, with R² metrics improved by 45% when employing feature selection for prediction. While ANN has been commonly used for solar generation prediction, this study uniquely combines ANN with PCC as a feature selection technique. This method has not been widely explored in the literature and has resulted in significant improvements in the accuracy and robustness of the prediction model. Thus, our study provides a valuable contribution to solar energy prediction and can potentially enhance solar energy utilisation efficiency in various applications. © 2023 The Authors