An online-learning-enabled self-attention-based model for ultra-short-term wind power forecasting

• Published in: Energy (Oxford) Vol. 272; p. 127173
• Main Authors: Dai, Xiaoran, Liu, Guo-Ping, Hu, Wenshan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.06.2023
• Subjects: Online learning; Self-attention-based model; Time series prediction; Wind power forecasting; Self-attention-based model; Time series prediction; Wind power forecasting; Online learning
• ISSN: 0360-5442
• DOI: 10.1016/j.energy.2023.127173

Renewable wind power accounts for an increasing proportion of the smart grid nowadays. The intermittent and fluctuating nature of wind renders wind power forecasting important. Recently, deep learning techniques have shown great potential in wind power forecasting, yet the existing methods mainly employ the “offline training–online forecasting” scheme which cannot capture the time-varying relations in wind power sequences. In this paper, a self-attention-based neural network (SANN) is conceived for online learning. Explicitly, the SANN model captures the temporal relations in power sequences via the self-attention mechanism. Unlike the popular recurrent deep learning structure in time series prediction, the SANN model is recurrence-free and allows parallel computation in the procedure of online learning. Meanwhile, the online learning algorithm is capable of adapting to the weather, operational, and several environmental variations, thus improving the forecasting accuracy. Finally, experiments are carried out on two real-world datasets with different characteristics. The experiments assure that our approach is superior to the conventional counterpart and hence validate the effectiveness. •The online learning strategy is designed for wind power forecasting, even with the time-varying time series pattern.•The self-attention mechanism is improved to capture the temporal relationship for increasing forecasting accuracy.•A cost-efficient model is developed specifically for online learning which has less computational complexity than recurrence-based models.