Research map / Time series / Long-term forecasting
U-Mixer: An Unet-Mixer Architecture with Stationarity Correction for Time Series Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes U-Mixer, a Unet-style MLP-Mixer for time series forecasting with a stationarity correction that restores non-stationary information via autocorrelation-matrix-based affine rescaling.
Problem
Non-stationarity in time series degrades deep forecasting models: it disrupts feature propagation, shifts feature distributions across channels, and makes it hard for models to learn distribution changes after stationarization.
Approach
Combines a multi-level Unet encoder-decoder with Mixer-style MLP blocks that separately mix temporal (per-channel) and channel dimensions, plus a stationarity correction that rescales the model output so its autocorrelation matrix matches the input's (computed via FFT/Wiener-Khinchin) and restores the mean difference.
Key results
- Achieves best MSE in 56 of 64 long-term forecasting settings across 8 datasets, with 14.5% MSE and 7.7% MAE improvements over the best existing results (per table improvement row vs TimesNet).
- On M4 short-term forecasting, averaged OWA 0.845 and SMAPE 11.740, slightly better than TimesNet (OWA 0.851).
- Ablation on ETTh2, Weather and M4 shows removing the Unet encoder-decoder or stationarity correction worsens MSE/MAE (e.g., ETTh2 MSE 0.381 vs 0.405 w/o UE and 0.400 w/o SC).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use U-Mixer instead of TimesNet | long-term forecasting · ETTh2 (Avg) | MSE: 0.414 -> 0.381 | Helped |
| Use U-Mixer instead of TimesNet | long-term forecasting · Weather (Avg) | MAE: 0.287 -> 0.260 | Helped |
| Add the Unet encoder-decoder (ablation) | U-Mixer · ETTh2 | MSE: 0.405 -> 0.381 | Helped |
| Add the Unet encoder-decoder (ablation) | U-Mixer · Weather | MAE: 0.285 -> 0.260 | Helped |
| Add stationarity correction (ablation) | U-Mixer · ETTh2 | MSE: 0.400 -> 0.381 | Helped |
| Add stationarity correction (ablation) | U-Mixer · M4 | MASE: 1.587 -> 1.575 | Helped |
Surprising findings
- Basic linear models (DLinear) are competitive with complex Transformers on long-term benchmarks, motivating the MLP-based design.
Limitations
- Sensitivity to patch length P is notably higher on ETTm2 than on Exchange.
- Stationarity correction relies on a heuristic affine rescaling; the approach is not derived for general distribution shifts.
- Results on ETTm2 at longer horizons (336, 720) are not best in all cases, with TimesNet/ETSformer winning some settings.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Traffic Exchange Weather M4
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Electricity Traffic Exchange Weather M4
Metrics: MSE MAE SMAPE MASE OWA
Baselines: TimesNet ETSformer LightTS DLinear FEDformer Autoformer LSSL LSTM TCN LogTrans Reformer Informer
Builds on: MLP-Mixer Unet Non-stationary Transformer DLinear N-BEATS
Closest papers by meaning
- xLSTM-Mixer: Multivariate Time Series Forecasting by Mixing via Scalar Memories (NeurIPS 2025 · similarity 0.89)
- Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series (NeurIPS 2024 · similarity 0.86)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.85)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.85)
- WPMixer: Efficient Multi-Resolution Mixing for Long-Term Time Series Forecasting (AAAI 2025 · similarity 0.85)
- Unified Training of Universal Time Series Forecasting Transformers (ICML 2024 · similarity 0.85)
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.