$K^2$VAE: A Koopman-Kalman Enhanced Variational AutoEncoder for Probabilistic Time Series Forecasting

The paper introduces $K^2$VAE, a VAE-based generative model that improves long-term probabilistic time series forecasting by transforming nonlinear dynamics into a linear system using KoopmanNet and refining predictions with KalmanNet, thereby reducing error accumulation. Extensive experiments show that $K^2$VAE outperforms existing methods in both short- and long-term forecasting, offering a more efficient and accurate approach.