The accuracy improvement of ultra-short-term wind power forecast with deep learning methods is of great significance for intraday unit commitment, ultra-short-term economic dispatch, and reserve scheduling of the power systems, which can further enhance the safety and efficiency. To address the problem in the existing feature extraction models that the similarity of implicit features and changing trends in time series curves have not been adequately extracted, this article proposes an ultra-short-term wind power forecast model based on contrastive learning-assisted training, which mainly consists of an input encoding module, a feature extraction module, a contrastive learning module, and a regression module. The self-supervised contrastive learning module autonomously generates positive and negative samples and enlarges the distance between the positive and negative samples in the projection space, which help to extract the implicit features of the similarity of the input information. In this way, the redundant information is reduced, the sample correlation is enhanced, and the accuracy of wind power forecast is ultimately improved. Compared with LSTM and Lightgbm methods, experimental results show that the mean absolute error of the proposed method is decreased by 19. 9% and 6. 5% , which effectively increase the wind power prediction accuracy.