Incremental transfer learning based on temporal-frequency convolution interaction for multi-task prediction of wind...

Area of Science:

• Renewable Energy
• Machine Learning
• Time Series Analysis

Background:

• Wind speed and power prediction are crucial for wind energy but challenged by data heterogeneity and time-varying properties.
• Existing models struggle with new wind farms due to limited historical data and data confidentiality issues.

Purpose of the Study:

• To develop an accurate and efficient multi-task prediction model for wind speed and power.
• To address the challenge of limited historical data in new wind farms using incremental transfer learning.

Main Methods:

• An incremental transfer learning framework was proposed for simultaneous wind speed and power prediction.
• A temporal-frequency convolutional interactive neural network integrated with circular convolution and a gated recurrent unit was developed.
• The model refines historical predictions with new data to capture evolving statistical characteristics.

Main Results:

• The proposed model achieved the highest prediction accuracy on real-world data from Northern China.
• The incremental transfer learning approach demonstrated enhanced predictive performance over extended periods.
• The model effectively captures dynamic behavior from limited samples, reducing computational requirements.

Conclusions:

• The developed model shows significant potential for practical engineering applications in wind energy.
• Incremental transfer learning is effective in improving prediction accuracy for evolving wind data patterns.
• The study highlights a viable solution for accurate wind power forecasting with limited data.