Efficient 24-hour ahead PV energy production forecasting employing a transformer-based model

• Published in: 2022 2nd International Conference on Energy Transition in the Mediterranean Area (SyNERGY MED) pp. 1 - 6
• Main Authors: Kothona, Despoina, Spyropoulos, Konstantinos, Valelis, Christos, Sarigiannidis, Charalampos, Chatzisavvas, Konstantinos Ch, Christoforidis, Georgios C.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.10.2022
• Subjects: Convolutional Neural Network; Data models; Deep Learning; Input variables; Long Short-Term Memory; Photovoltaic power forecasting; Photovoltaic systems; Planning; Predictive models; Production; Transformer; Transformers
• DOI: 10.1109/SyNERGYMED55767.2022.9941461

The increased installation of photovoltaic (PV) systems has triggered several issues in power systems, due to their intermittent nature. Therefore, the development of accurate day-ahead PV generation forecasting models is vital to overcome planning and dispatching issues. The present paper presents a day-ahead PV energy forecasting model based on Transformer networks. The model is executed in hourly basis and provides 24-hours ahead forecasts. The accuracy of the proposed model is compared to three Deep Learning (DL) methods, i.e., Long Short-Term Memory (LSTM), Bidirectional Stacked LSTM (BiSLSTM) and Convolutional Neural Network (CNN). Additionally, three scenarios have been implemented to examine the forecasting error in terms of the input variables. All models and scenarios are tested on data from five PV plants. The results highlight that the Transformer-based model outperforms the other DL models, since it increases R-squared (R 2 ) up to 14.1% and reduces the Mean Absolute Error (MAE) up to 27.9%.