Research map / Time series / Long-term forecasting
DDN: Dual-domain Dynamic Normalization for Non-stationary 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 DDN, a model-agnostic plug-in that performs sliding-window normalization in both wavelet frequency and time domains and reconstructs non-stationary statistics via a distribution prediction model for time series forecasting.
Problem
DNN-based time series forecasters give unreliable predictions on non-stationary data whose distribution drifts over time; existing normalization methods use fixed windows or slices and struggle to capture rapid, time-varying distribution changes.
Approach
Combine point-level sliding-window normalization in the time domain with sliding normalization of DWT low- and high-frequency components (learnable wavelet bases), weight the two stationary outputs, and reconstruct future mean/std via a distribution prediction model trained with pretraining then collaborative training.
Key results
- DDN applied to Autoformer reduces MSE by 19.2%, 24.7%, and 25.6% on Weather, Electricity, and Traffic respectively.
- DDN applied to FEDformer reduces MSE by 13.1%, 16.2%, and 22.3% on Weather, Electricity, and Traffic respectively.
- On Traffic, SAN with DLinear/iTransformer performs worse than the base model, while DDN improves it (e.g., iTransformer+DDN MSE 0.336 vs 0.348 at horizon 96).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DDN normalization | Autoformer · Weather, Electricity, Traffic (96-step forecast) | MSE: 0.247 -> 0.190; 0.195 -> 0.150; 0.654 -> 0.453 | Helped |
| Add DDN normalization | FEDformer · Weather, Electricity, Traffic (96-step forecast) | MSE: 0.246 -> 0.174; 0.185 -> 0.146; 0.579 -> 0.442 | Helped |
| Add DDN normalization | DLinear · Weather, Electricity, Traffic (96-step forecast) | MSE: 0.175 -> 0.146; 0.140 -> 0.131; 0.411 -> 0.375 | Helped |
| Replace RevIN with DDN | iTransformer · Weather, Electricity, Traffic (96-step forecast) | MSE: 0.177 -> 0.148; 0.133 -> 0.127; 0.348 -> 0.336 | Helped |
| Add time-domain normalization (TDN) alongside frequency-domain normalization (FDN) (ablation) | DLinear + FDN Only · Weather | MSE: 0.149 -> 0.146 | Helped |
| Add frequency-domain normalization (FDN) alongside time-domain normalization (TDN) (ablation) | DLinear + TDN Only · Weather | MSE: 0.150 -> 0.146 | Helped |
| Add time-domain normalization (TDN) alongside frequency-domain normalization (FDN) (ablation) | DLinear + FDN Only · Electricity | MSE: 0.132 -> 0.131 | Helped |
| Add frequency-domain normalization (FDN) alongside time-domain normalization (TDN) (ablation) | DLinear + TDN Only · Electricity | MSE: 0.133 -> 0.131 | Helped |
| Add time-domain normalization (TDN) alongside frequency-domain normalization (FDN) (ablation) | DLinear + FDN Only · Traffic | MSE: 0.375 -> 0.375 | No clear effect |
| Add frequency-domain normalization (FDN) alongside time-domain normalization (TDN) (ablation) | DLinear + TDN Only · Traffic | MSE: 0.378 -> 0.375 | Helped |
| Add time-domain normalization (TDN) alongside frequency-domain normalization (FDN) (ablation) | iTransformer + FDN Only · Electricity | MSE: 0.126 -> 0.127 | Hurt |
| Add both time- and frequency-domain normalization (ablation) | DLinear · Weather, Electricity, Traffic | MSE: Table 5 | Mixed |
Surprising findings
- SAN degrades performance relative to the original predictive model on the Traffic dataset, according to the authors.
Limitations
- Ablation shows FDN-only and TDN-only sometimes match DDN, so gains from combining domains are modest in some settings.
- Distribution ground truth is derived from an assumption-based sliding statistic, which the authors acknowledge is inconsistent with actual high-frequency changes.
Experimental setup
Datasets: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic
Benchmarks: ETTh1 ETTh2 ETTm1 ETTm2 Weather Electricity Traffic
Metrics: MSE MAE
Baselines: RevIN NST Dish-TS SAN Autoformer FEDformer DLinear iTransformer
Builds on: RevIN SAN Dish-TS Non-stationary Transformers iTransformer
Closest papers by meaning
- Enhancing Foundation Models for Time Series Forecasting via Wavelet-based Tokenization (ICML 2025 · similarity 0.88)
- SIN: Selective and Interpretable Normalization for Long-Term Time Series Forecasting (ICML 2024 · similarity 0.87)
- TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.87)
- FilterNet: Harnessing Frequency Filters for Time Series Forecasting (NeurIPS 2024 · similarity 0.86)
- Hierarchical Classification Auxiliary Network for Time Series Forecasting (AAAI 2025 · similarity 0.86)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.86)
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.