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

2026-06-25 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

PMDformer is a novel Transformer-based model designed for long-term time series forecasting (LTSF), addressing challenges in accurately modeling shape similarities across data patches and variables due to scale differences. Introduced on 2026-06-25, PMDformer incorporates Patch-Mean Decoupling (PMD), which separates trend and residual shape information by subtracting the mean of each patch, thereby preserving original data structure for attention mechanisms. The model further enhances long-range dependency and cross-variable relationship modeling through two key components: Trend Restoration Attention (TRA), which reintegrates the decoupled trend during attention output calculation, and Proximal Variable Attention (PVA), which focuses cross-variable attention on recent, relevant time segments to avoid overfitting. Extensive experiments demonstrate that PMDformer achieves superior stability and accuracy compared to existing leading methods across multiple LTSF benchmarks.

Key takeaway

For Machine Learning Engineers developing long-term time series forecasting solutions, PMDformer offers a robust approach to overcome challenges with scale differences and shape similarity. You should consider implementing patch-mean decoupling and focused cross-variable attention, as demonstrated by PMDformer, to enhance model stability and accuracy. This method can significantly improve your model's ability to capture true dependencies without overfitting on outdated correlations.

Key insights

Decoupling trend and residual shape information enhances Transformer attention for long-term time series forecasting.

Principles

• Decouple trend from shape to improve attention mechanism accuracy.
• Reintegrate trend information during attention output calculation.
• Focus cross-variable attention on recent data to mitigate overfitting.

Method

PMDformer employs Patch-Mean Decoupling (PMD) to separate trend and shape, then uses Trend Restoration Attention (TRA) and Proximal Variable Attention (PVA) for robust long-range and cross-variable dependency modeling.

In practice

• Apply patch-mean decoupling for time series shape analysis
• Utilize recent data focus for cross-variable attention

Topics

• Long-term Time Series Forecasting
• Transformer Models
• Patch-Mean Decoupling
• Attention Mechanisms
• Cross-Variable Relationships
• Time Series Analysis

Code references

• aohu1105/PMDformer

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Causal Semantic Alignment for LLM-based Time Series Forecasting — CVAformer disentangles time series variables into invariant and dynamic components to improve LLM-based forecasting accuracy by mitigating confounding effects.
• From Long News to Accurate Forecast: Importance-Aware Fusion and PRM-Guided Reflection for Time Series Forecasting — Importance-aware compression and PRM-guided retrieval improve news-enhanced time series forecasting by managing context and selection.
• XCTFormer: Leveraging Cross-Channel and Cross-Time Dependencies for Enhanced Time-Series Analysis — XCTFormer explicitly models cross-temporal and cross-channel dependencies in multivariate time-series using an enhanced transformer attention mechanism.
• Phys-JEPA: Physics-Informed Latent World Models for Multivariate Time-Series Forecasting — Phys-JEPA improves multivariate time-series forecasting by imposing physical consistency directly on latent states and transitions.
• Forecast collapse of transformer-based models under squared loss in financial time series — Highly expressive models like Transformers degrade on financial time series due to increased variance from noise reuse.
• Forecasting as Rendering: A 2D Gaussian Splatting Framework for Time Series Forecasting — Time series forecasting can be fundamentally improved by treating it as 2D generative rendering using anisotropic Gaussian splatting.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.