Latent adversarial regularized autoencoder for high-dimensional probabilistic time series prediction

Jing Zhang, Qun Dai

Abstract: Highlights•TimeLAR is a Transformer-based autoencoder model combined with GANs for TSP.•TimeLAR leverages autoencoder to provide a lower-dimensional latent space.•A modified Transformer is designed to infer latent space prediction distributions.•TimeLAR employs GANs to refine the performance of latent space predictions.•TimeLAR is trained with reconstruction, prediction, and adversarial losses.