DiffAT: Effective data augmentation with diffusion models for time series forecasting

Data augmentation offers a promising solution to data scarcity in deep learning-based time series forecasting. However, current approaches face dual limitations (1) Hand-designed methods (e. g. , cropping/masking): often disrupt the continuity of vital temporal patterns (such as seasonal and trend) by introducing abrupt pattern discontinuities; (2) Generative models often face difficulties in preserving task-critical features that are essential for prediction, especially when aiming to generate diverse augmented series. To tackle these dilemmas, we propose a novel conditional diffusion-based data augmentation framework, named DiffAT, for time series forecasting tasks. DiffAT synergizes: (1) Patch-wise masking reconstruction to capture structural invariants (such as autocorrelation and causality), and (2) encoding hand-designed augmentation prototypes for guiding diversity-preserving generation. DiffAT achieves dual enhancement: maintaining continuity of temporal patterns through progressive denoising process and exposing latent invariant patterns via guided diversity injection. We validate the efficacy of DiffAT through extensive experiments on seven real-world datasets, by comparing DiffAT with six state-of-the-art time series data augmentation methods. The results indicate our method can boost the forecasting performance of Autoformer by up to 6. 49 % in 26/28 cases and improve forecasting performance of LightTS by up to 3. 11 % in 23/28 cases on 7 real-world benchmarks. Extensive experiments also indicate that DiffAT can improve the accuracy of forecasting models in few-shot scenario (with 1 % training data) in 54/60 cases. We will release the source code upon publication.

Authors

• Yang Yu Nanjing University
• Ruizhe Ma
• Wenbo Gu
• Zongmin Ma

Keywords

• Data augmentation
• Time series forecasting
• Diffusion model
• Classifier-guidance generation