Abstract: To overcome limitations in existing self-supervised methods for time series, the Soft Whitening inspired non-contrastive self-supervised learning framework for Time-series Forecasting (SWTF) was proposed. Key limitations addressed were the reliance on negative pairs and the disruption of temporal dependencies by data augmentation. The SWTF framework, a negative-sample-free and self-supervised approach, was designed to learn information-dense and predictive representations. Its core methodology was based on two collaboratively optimized objectives. The first was a Temporal Soft Whitening (TSW) loss, which learned compact representations by reducing feature redundancy while preserving variance. This method eliminated the need for negative pairs and representation invariance constraints. The second was a denoising reconstruction task designed to guide the model in encoding predictive temporal features. To jointly optimize these objectives, a Siamese encoder with parallel multi-scale convolutional networks was constructed. This architecture efficiently captured both micro-level local periodicities and macro-level long-range dependencies. On several benchmark datasets (electricity, weather, and traffic), SWTF achieves a Mean Absolute Error (MAE) reduction of approximately 19.6% (13.4% to 26.4%) compared to mainstream baselines like CoST and PDF. These results demonstrate the effectiveness of the SWTF framework in learning robust and highly predictive time series representations.