Short-term multi-step wind power prediction model based on Pt-Transformer neural network integrating spatio-temporal feature and sparse attention

•A wind power forecasting method based on the time series Transformer model is proposed.•Novel model improves noisy data handling, time-series dependency capture, and efficiency.•Introduce a multi-step autoregressive prediction framework using Teacher forcing. The accuracy and stability of wind powe...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Electric power systems research Jg. 248; S. 111970
Hauptverfasser: Zhang, Zhiyan, Sun, Zhenyang, Guo, Xianghui, Guo, Ruipeng, Yang, Xiaoliang, Yang, Pengju
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier B.V 01.11.2025
Schlagworte:
Seasonal trend decomposition
Sparse attention mechanism
Temporal autoencoder
Transformer
Wind power prediction
Seasonal trend decomposition
Sparse attention mechanism
Wind power prediction
Transformer
Temporal autoencoder
ISSN:0378-7796
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:•A wind power forecasting method based on the time series Transformer model is proposed.•Novel model improves noisy data handling, time-series dependency capture, and efficiency.•Introduce a multi-step autoregressive prediction framework using Teacher forcing. The accuracy and stability of wind power prediction are crucial for grid dispatching. However, precise wind power prediction faces three major challenges: effective mitigating data noise, representing complex spatio-temporal features, and selecting an appropriate prediction model. To address these challenges, this paper proposes a new wind power prediction model called process temporal transformer (Pt-Transformer). Firstly, an exponential weighted moving average (EWMA) method is employed to filter the noise of original wind characteristics while maintaining data complexity. Secondly, a seasonal trend decomposition (STD) model combined with a temporal autoencoder (TAE) is implemented to efficiently extract complex spatio-temporal characteristics. Furthermore, a sparse attention mechanism combined with the Transformer architecture is introduced to effectively extract high-dimensional latent space representations. Finally, two case studies are conducted using data from a cluster of wind farms and a single wind farm in central China, and three models, namely Transformer, Prob-Transformer, and Pt-Transformer, are used for prediction. The results show that the Pt-Transformer model outperforms the other two models in prediction accuracy and stability, reflecting its excellent performance in short-term multi-step wind power generation prediction.
ISSN:0378-7796
DOI:10.1016/j.epsr.2025.111970