Wind power prediction using stacking and transfer learning.

Area of Science:

• Renewable Energy
• Electrical Engineering
• Data Science

Background:

• Accurate wind power forecasting is essential for grid stability and economic efficiency as renewable energy adoption increases.
• Existing prediction methods face challenges in achieving high accuracy for ultra-short-term wind power output.

Purpose of the Study:

• To develop and evaluate a novel ensemble method for enhancing ultra-short-term wind power prediction accuracy.
• To leverage Stacking and Transfer Learning techniques to improve forecasting performance.

Main Methods:

• Data dimensionality reduction using Principal Component Analysis (PCA).
• Utilizing Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), Gated Recurrent Unit (GRU), Bidirectional GRU (BiGRU), and LSTM-Attention as base predictive models.
• Employing a Stacking ensemble to combine predictions from base models.
• Implementing Transfer Learning to enhance model generalization and performance.

Main Results:

• The proposed Stacking and Transfer Learning method demonstrated superior accuracy in ultra-short-term wind power prediction compared to individual base models.
• PCA effectively reduced data dimensions, aiding in model efficiency.
• The ensemble approach successfully integrated diverse deep learning models for robust forecasting.

Conclusions:

• The combined Stacking and Transfer Learning approach offers a significant advancement in ultra-short-term wind power forecasting.
• This method provides a more reliable tool for grid operators to manage fluctuating wind energy resources.
• Future work could explore additional feature engineering and hyperparameter optimization for further improvements.