SRT: Super-Resolution for Time Series via Disentangled Rectified Flow

The article introduces SRT (Super-Resolution for Time Series), a novel framework designed to reconstruct high-resolution time series data from low-resolution inputs using disentangled rectified flow. SRT decomposes inputs into trend and seasonal components and utilizes an implicit neural representation along with a cross-resolution attention mechanism, while the larger variant, SRT-large, shows strong zero-shot super-resolution capabilities after extensive pre-training. Experiments across nine public datasets indicate that both SRT and SRT-large outperform existing methods, highlighting their practical relevance for practitioners needing high-quality temporal data reconstruction.