FRIREN introduces a new approach to long-term time-series forecasting using geometric structure as the foundational model.
The model minimizes Wasserstein-2 distance and provides a spectral view of dynamics for forecasting.
FRIREN utilizes Flow-inspired Representations via Interpretable Eigen-networks for prediction, achieving superior results on chaotic systems like Lorenz-63 and Rossler compared to TimeMixer.
The model showcases high accuracy and interpretability in long-term forecasting, setting a new benchmark in LTSF model design.