CPDD: Generalized Compressed Representation for Multivariate Long-term Time Series Generation
Keywords: Generative Model, Deep Learning, Mode Function, Diffusion Model, Long-term Time Series
TL;DR: We develop a novel generative framework, CPDD, specifically for long-term time series generation, exploring the cross-scale feature fusion based on the patch mode decomposition method.
Abstract: The generation of time series has increasingly wide applications in many fields, such as electricity and energy. Generating realistic multivariate long time series is a crucial step towards making time series generative models practical, with the challenge being the balance between long-term dependencies and short-term feature learning. Towards this end, we propose a novel time series generative model named Compressed Patch Denoising Diffusion-model (CPDD). Concretely, CPDD first employs the Time-series Patch Compressed (TPC) module based on the patch mode decomposition method to obtain the latent encoding of multi-scale feature fusion. Subsequently, it utilizes a diffusion-based model to learn the latent distribution and decode the resulting samples, thereby achieving high-quality multivariate long-time series generation. Through extensive experiments, results show that CPDD achieves state-of-the-art performance in the generation task of multivariate long-time series. Furthermore, TPC also exhibits remarkable efficiency in terms of robustness and generalization in time series reconstruction.
Supplementary Material: zip
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 10767
Loading