Data-augmented trend-fluctuation representations by interpretable contrastive learning for wind power forecasting

Data augmentation can expand wind power data by analyzing their statistical characteristics, providing richer input information for forecasting models, thereby improving the forecasting accuracy. However, existing data augmentation methods only learn the probability distribution of original data, making it difficult for them to capture and represent complex trend and fluctuation features from data. Additionally, heterogeneous data patterns from different wind farms affect the generalization of forecasting models and the black-box structure of deep learning models is not trustworthy in practical applications. Therefore, a novel interpretable contrastive learning framework of trend-fluctuation representations (ICoTF) is proposed for wind power forecasting. Specifically, ICoTF includes a pretraining stage and a regression stage. Initially, data augmentation based on contrastive pretraining is designed to extract trend and fluctuation representations from wind power data, assisted by a time-frequency domain contrastive loss. In the regression stage, these representations are fed into a personalized ridge regression model, and its parameters are fine-tuned by mean squared error (MSE) loss to achieve high-performance forecasting. Furthermore, an optimal transport algorithm is integrated into the contrastive loss to reveal the interactions between various input features and the importance of each feature to wind power forecasts, thus achieving interpretable learning. The proposed model is evaluated on two datasets, and the results demonstrate that ICoTF exhibits superior forecasting accuracy, generalization ability and interpretability compared to other benchmark models.