TriTS Framework Introduces Cross-Modal Approach for Long-Term Time Series Forecasting

The TriTS framework introduces an innovative solution to the ongoing issue of Long-term Time Series Forecasting (LTSF) by transforming one-dimensional time series into orthogonal spaces of time, frequency, and two-dimensional vision. This method of cross-modal disentanglement seeks to effectively represent complex temporal dynamics that are challenging to capture in a solely 1D format. To connect the 1D and 2D modalities without incurring the quadratic computational costs associated with Vision Transformers, it employs a Period-Aware Reshaping strategy alongside Visual Mamba (Vim). This synergy captures cross-period dependencies as global visual textures while ensuring linear computational efficiency. Additionally, a Multi-Resolution Wavelet Mixing component enhances the framework. The significance of time series forecasting spans finance, energy, transportation, and meteorology. This research was shared on arXiv with the identifier 2604.16748v1 and categorized as a cross announcement.