LFformer: An improved Transformer model for wind power prediction.

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence in Energy
• Time Series Forecasting

Background:

• Wind power forecasting faces challenges due to complex nonlinearities and multi-scale temporal patterns.
• Traditional forecasting methods struggle to accurately capture these intricate dynamics, impacting grid stability and operational efficiency.

Purpose of the Study:

• To develop an innovative ultra-short-term wind power forecasting model, LFformer, capable of addressing multi-scale complexities.
• To enhance the accuracy and stability of wind power predictions, particularly in the crucial ultra-short-term horizon.

Main Methods:

• The LFformer model utilizes an encoder-decoder architecture for information focus and Devlin normalization for data scaling.
• Legendre polynomials project data into a bounded space, followed by feature compression and selection via Fourier Transform's low-rank approximation.
• A multi-scale mixing mechanism feeds predictions into a multilayer perceptron for final output after back-normalization.

Main Results:

• The LFformer model demonstrated superior prediction accuracy and stability compared to existing methods.
• Significant advantages were observed in ultra-short-term wind power forecasting scenarios.
• The model effectively handles complex nonlinear features and behavioral patterns across various time scales.

Conclusions:

• The proposed LFformer model offers a robust solution for ultra-short-term wind power forecasting.
• Improved forecasting enhances overall wind power system efficiency and power grid stability.
• This research provides technical support for wind power enterprises in competitive energy markets.