A Self-Supervised Pre-Learning Method for Low Wind Power Forecasting

As wind power is becoming a major energy source of power systems, the risk of power shortages due to its intermittent low power output is growing. Accurate forecasting of low wind power is crucial for mitigating these impacts. However, conventional methods struggle with few-sample issues due to the infrequent occurrence of low wind power, limiting accuracy improvements. To address this, a self-supervised pre-learning method is proposed to forecast low wind power occurrence period and output, leveraging the similarities and differences among low output samples to enhance forecasting accuracy. Low wind power output is decomposed into low wind power events (LWPE), which represent the occurrence timeframe, and low wind power processes (LWPP), which represent the power sequences. For LWPE forecasting, a siamese residual shrinkage network based on contrastive learning is introduced. This network pre-learns LWPE features from sample pairs to mitigate the impact of imbalanced sample distribution. For LWPP forecasting, a pattern recognition-based embedded forecasting framework is proposed, embedding typical LWPP fluctuations into the prediction network to improve fit under limited sample conditions. A case study on 3 wind farm clusters shows that this method improves LWPP forecasting accuracy from 84.99%-86.6% to 89.97%, outperforming traditional methods without pre-learning.