Research map / Time series / Long-term forecasting
PMDformer: Patch-Mean Decoupling Transformer for Long-term Forecasting
Ao Hu, Liangjian Wen, Jiang Duan, Yong Dai, YAN HE, Dongkai Wang, Jun Wang, Yukun Zhang, Ruoxi Jiang, Zenglin Xu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use patch-mean decoupling instead of mean-variance patch normalization (ablation) | PMDformer · Traffic | MSE: 0.396 -> 0.378 | Improved |
| use patch-mean decoupling instead of SAN (ablation) | PMDformer · ETTh2 | MSE: 0.360 -> 0.337 | Improved |
| add patch-mean decoupling versus no decoupling (ablation) | PMDformer · Solar | MSE: 0.199 -> 0.181 | Improved |
| use patch-mean decoupling instead of mean-variance normalization (ablation) | PMDformer · ETTm1 and Weather | MSE: ETTm1 0.347 -> 0.346; Weather 0.218 -> 0.217 | No clear effect |
| restrict cross-variable attention to the last patch (PVA) (ablation) | PMDformer · Solar | MSE: 0.186 -> 0.181 | Improved |
| add PVA to model without PVA (ablation) | PMDformer · Traffic | MSE: 0.386 -> 0.378 | Improved |
| use trend restoration attention instead of standard self-attention (ablation) | PMDformer · Traffic | MSE: 0.388 -> 0.378 | Improved |
| add TRA and PVA together (ablation) | PMDformer · Traffic | MSE: 0.426 -> 0.378 | Improved |
| apply PVA before TRA instead of TRA before PVA (ablation) | PMDformer · Solar | MSE: 0.188 (swapped) -> 0.181 | Improved |
| use cross-variable attention over k nearest patches larger than 1 (ablation) | PMDformer · ETTh1 and Solar | MSE | Hurt |
| use moderate patch sizes 24 to 72 instead of very small or large (ablation) | PMDformer · ETTh2 and ECL | MSE | Improved |
| use PMDformer instead of TimeBase | TimeBase · eight LTSF datasets | MSE: 5.68% average MSE reduction | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TimeBase — TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025) · better on 5 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Unlocking the Power of Patch: Patch-Based MLP for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.91)
- DeformableTST: Transformer for Time Series Forecasting without Over-reliance on Patching (NeurIPS 2024 · similarity 0.89)
- A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle (ICML 2025 · similarity 0.87)
- Linear Transformers as VAR Models: Aligning Autoregressive Attention Mechanisms with Autoregressive Forecasting (ICML 2025 · similarity 0.87)
- iTransformer: Inverted Transformers Are Effective for Time Series Forecasting (ICLR 2024 · similarity 0.87)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
