Research map / Time series / Long-term forecasting
PGN: The RNN's New Successor is Effective for Long-Range 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 Parallel Gated Network (PGN), an RNN successor with O(1) propagation paths via a linear historical-information layer plus a single gate, and TPGN, a 2D two-branch temporal framework for long-range forecasting that reports SOTA on five benchmarks.
Problem
RNNs have long information propagation paths, causing difficulty capturing long-term dependencies, gradient explosion/vanishing, and slow sequential execution; existing time series models also struggle to directly capture periodicity when modeling in 1D.
Approach
Replace the recurrent structure with a parallel linear historical-information extraction layer over all prior steps, fused with the current input by a single gate, and apply it to 2D-reshaped series (PGN on the column/period dimension) alongside a patch-based short-term branch.
Key results
- TPGN achieves best MSE in all 20 long-range settings (input 168, horizons 168-1440) on five datasets, with average MSE improvement of 12.35% over previous best methods.
- Average MSE improvement of TPGN over individual baselines ranges from 14.08% (WITRAN) to 37.44% (PatchTST).
- In ablation, TPGN-GRU/LSTM/MLP variants are worse than TPGN with PGN, and single-branch variants degrade performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use both long-term and short-term branches instead of only the long-term branch (ablation) | TPGN · Long-range forecasting; five datasets | MSE: 0.2710 -> 0.2484 | Helped |
| Add the long-term branch to the short-term-only model (ablation) | TPGN · Long-range forecasting; five datasets | MSE: 0.8323 -> 0.2484 | Helped |
| Use PGN instead of GRU in the long-term branch (ablation) | TPGN · Long-range forecasting; five datasets | MSE: 0.3557 -> 0.2484 | Helped |
| Use PGN instead of LSTM in the long-term branch (ablation) | TPGN · Long-range forecasting; five datasets | MSE: 0.2936 -> 0.2484 | Helped |
| Use PGN instead of MLP in the long-term branch (ablation) | TPGN · Long-range forecasting; five datasets | MSE: 0.3456 -> 0.2484 | Helped |
Surprising findings
- Using only the short-term patch branch (TPGN-short) yields much worse results than the full model, and the long-term PGN branch alone is nearly as strong, especially on Traffic.
Limitations
- Focuses on temporal modeling only and does not model inter-variable dependencies; experiments use single-variable setup.
- TPGN does not have the lowest time and memory overhead among compared models.
- Hyperparameter normalization choice (norm) must be set per dataset.
Experimental setup
Datasets: ECL Traffic ETTh1 ETTh2 Weather
Benchmarks: ECL Traffic ETTh1 ETTh2 Weather
Metrics: MSE MAE
Baselines: WITRAN ModernTCN TimesNet MICN FITS TimeMixer DLinear iTransformer PDF Basisformer PatchTST FiLM
Builds on: WITRAN TimesNet PatchTST GRU LSTM
Closest papers by meaning
- TempoPFN: Synthetic Pre-training of Linear RNNs for Zero-shot Time Series Forecasting (NeurIPS 2026 · similarity 0.87)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.87)
- TStruct: Learning Shared Temporal Structures for Long-Term Time-Series Forecasting (NeurIPS 2026 · similarity 0.86)
- TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data (AAAI 2025 · similarity 0.86)
- CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns (NeurIPS 2024 · similarity 0.86)
- Unlocking the Power of LSTM for Long Term Time Series Forecasting (AAAI 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.