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

2026-06-24

AI Summary

This work addresses the challenge posed by scale discrepancies in long-term time series forecasting, which hinder accurate modeling of shape similarities across segments and variables. To overcome this, the authors propose a novel Transformer-based architecture featuring three key innovations: Patch Mean Decomposition (PMD) to decouple trend and residual components while preserving original structural information; Trend Restoration Attention (TRA) to reconstruct global trends effectively; and Proximal Variable Attention (PVA) to focus on recently relevant variables, thereby avoiding spurious correlations from outdated associations. The proposed method consistently outperforms state-of-the-art models across multiple standard benchmarks, achieving significant and robust improvements in both prediction accuracy and stability.

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. 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 attention mechanism captures true shape similarities. Futhermore, to more effectively model long-range dependencies and capture cross-variable relationships, we propose Trend Restoration Attention (TRA) and Proximal Variable Attention (PVA). The former module reintegrates the decoupled trend from PMD while calculating attention output. And the latter focuses cross-variable attention on the most relevant, recent time segments to avoid overfitting on outdated correlations. Combining these components, we propose PMDformer, a model designed to effectively capture shape similarity in long-term forecasting scenarios. Extensive experiments indicate that PMDformer outperforms existing state-of-the-art methods in stability and accuracy across multiple LTSF benchmarks. The code is available at https://github.com/aohu1105/PMDformer.

Problem

• long-term time series forecasting
• shape similarity
• scale differences
• patch-based modeling
• cross-variable relationships

Innovation

• Patch-Mean Decoupling
• Trend Restoration Attention
• Proximal Variable Attention
• Shape Similarity
• Long-term Time Series Forecasting