AdaTS introduces a dynamic soft contrastive learning framework for adaptive time series representations, leveraging frequency-domain similarities with dynamic temporal and ordinal instance contrasts.