A Short-Term Wind Power Forecasting Method With Self-Supervised Contrastive Learning-Based Feature Extraction Model

It is significant and challenging to forecast wind power through the knowledge learned from the historical meteorological variables and wind power itself. So far, this knowledge has been learned based on supervised learning, which is insufficient in capturing semantic dependencies between variables and cannot provide a stability model. This work proposed an unsupervised representation learning-based short-term wind power prediction method using sequences contrastive technique. This representation learning model is mainly composed of the input encoder module, feature extraction module, contrastive learning module, and regression module. The feature extraction module which fuses multiple encoders can extract abundant feature vectors from the multi-source embedding of the input encoder module. The loss function in the contrastive learning module attempts to distinguish positive pairs and negative pairs that are generated from the feature extraction module. Eventually, the regression model uses the fully learned representation matrix for the power estimation. In order to explore the feasibility of this method in wind power prediction, the benefits under different prediction intervals are studied separately through some classic evaluation criteria, and comparing the proposed representation learning-based forecasting with other state-of-the-art methods. The results show that the knowledge learned from this model can generate improved performance with reliability and effectiveness with a 6-hour prediction horizon, and allow the development of short-term wind power forecast.