PMDformer: Patch-Mean Decoupling Transformer for Long-term Forecasting

ICLR 2026 Conference Submission19141 Authors

Keywords: Long-term Time Series Forecasting; Non-stationary; Patch-Mean Decoupled

Abstract: Long-term time series forecasting (LTSF) plays a crucial role in fields such as energy management, finance, and traffic prediction. Transformer-based models have adopted patch-based strategies to capture long-range dependencies, but accurately modeling shape similarities across patches and variables remains challenging due to scale differences. Traditional methods, like Patch Normalization, often distort shape information by removing amplitude variations. To address this, we introduce patch-mean decoupling (PMD), which separates the trend and residual shape information by subtracting the mean of each patch, preserving the original structure and ensuring that the model captures true shape similarities. Recognizing the importance of capturing shape relationships not just within patches but also across variables, we introduce Proximal Variable Attention (PVA), which focuses attention on the most relevant, recent time segments to avoid overfitting on outdated correlations. Finally, to restore global trend information without affecting shape similarity, we propose Trend Restoration Attention (TRA), which reintegrates the decoupled trend into the model’s attention mechanism. Combining these components, our model PMDformer outperforms existing state-of-the-art methods in stability and accuracy across multiple LTSF benchmarks.

Primary Area: learning on time series and dynamical systems

Submission Number: 19141