DDM-TS: Decoupled Diffusion Models for Time-Series Generation with Explicit Trend–Seasonality Decomposition

Jion Kim, Chanyoung Kim, Byeong-Seok Shin

Keywords: Diffusion Models, Time Series Generation, STL Decomposition, Multivariate Time Series, Unconditional Generation

TL;DR: We introduce DDM-TS, a decoupled diffusion framework that employs separate pathways for trend and seasonality modeling to generate high-fidelity, unconditional multivariate time series across extended temporal lengths.

Abstract: Time-series data are crucial for analysis and prediction across various domains, including finance, healthcare, and energy. However, issues of data scarcity, sensitivity concerns, and privacy regulations limit the availability of high-quality datasets, motivating growing research interest in time-series generation. While recent diffusion models have demonstrated stable learning capabilities and high-quality results for time-series generation, existing approaches face challenges in maintaining structural integrity when generating unconditional multivariate sequences over extended periods. This challenge involves preserving both long-term $\textit{trends}$ and periodic $\textit{seasonality}$ simultaneously, as unified frameworks suffer from inherent component interference that diminishes generation quality. To address this challenge, we propose DDM-TS, a novel decoupled diffusion framework that trains independent diffusion models for trend and seasonality components decomposed via STL decomposition. The framework then employs an adaptive gate-based fusion module that provides learnable fusion weights to unify the independently generated components into coherent synthetic time series. By decoupling trend and seasonal processing pathways, DDM-TS effectively alleviates mutual interference while preserving structural characteristics. Comprehensive experiments across benchmark datasets demonstrate that DDM-TS outperforms state-of-the-art baselines across diverse evaluation measures, achieving an average improvement of 33.8\% in reconstruction metrics (RMSE, PSNR), while preserving distributional coverage and frequency-domain consistency as demonstrated by t-SNE and spectral analysis metrics. Code is available at https://anonymous.4open.science/r/DDM-TS-0D44/.

Primary Area: learning on time series and dynamical systems

Submission Number: 6936