Research map / Time series / Long-term forecasting
TVNet: A Novel Time Series Analysis Method Based on Dynamic Convolution and 3D-Variation
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 TVNet, a purely convolutional time series model that reshapes 1D series into 3D tensors (inter-patch, intra-patch, cross-variable) and uses time-varying dynamic convolution weights, reporting SOTA across five time series tasks.
Problem
CNNs have lagged Transformer and MLP models for time series analysis, and existing convolutional time series models struggle to jointly capture intra-patch, inter-patch, and cross-variable dependencies.
Approach
A 3D-Embedding reshapes a patched, odd-even split 1D series into a 3D tensor, and a 3D-block generates time-varying convolution weights (a shared base weight scaled per patch by intra- and inter-patch pooled features) to model the three dependency types with dynamic 2D convolution.
Key results
- Long-term forecasting: TVNet achieves best average MSE on ETTm1 (0.348), ETTm2 (0.251), Weather (0.221) and ILI (1.406) among compared models, per the authors' tables.
- Short-term forecasting on M4: weighted-average SMAPE 11.671 and OWA 0.832, better than PatchTST (0.851) and U-Mixer (0.845).
- Classification average accuracy of 74.6% on 10 UEA datasets; anomaly detection reported as 86.8 (competitive with prior SOTA).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TVNet instead of the strongest listed baseline | Long-term forecasting · ETTm1, ETTm2, Weather, ILI | MSE: ETTm1 0.351 -> 0.348; ETTm2 0.255 -> 0.251; Weather 0.226 -> 0.221; ILI 1.443 -> 1.406 | Helped |
| Use TVNet instead of the strongest listed baseline | Long-term forecasting · ETTh1, ETTh2, Electricity, Traffic, Exchange | MSE: ETTh1 0.404 -> 0.407; ETTh2 0.320 -> 0.324; Electricity 0.156 -> 0.165; Traffic 0.391 -> 0.396; Exchange 0.297 -> 0.298 | Hurt |
| Use TVNet instead of the strongest listed baseline | Short-term forecasting · M4 WA | SMAPE, MASE, OWA: SMAPE 11.698 -> 11.671; MASE 1.575 -> 1.536; OWA 0.838 -> 0.832 | Helped |
| Use TVNet instead of the strongest listed baseline | Imputation · ETTm1, ETTh1, ETTh2, Weather | MSE: ETTm1 0.020 -> 0.018; ETTh1 0.050 -> 0.046; ETTh2 0.042 -> 0.039; Weather 0.027 -> 0.024 | Helped |
| Use TVNet instead of the strongest listed baseline | Imputation · ETTm2, Electricity | MSE: ETTm2 0.019 -> 0.022; Electricity 0.073 -> 0.079 | Hurt |
| Add inter-pool to TVNet (ablation) | TVNet · Weather, Electricity, Traffic; prediction lengths 96, 192, 336, 720 | MSE: Weather 0.156 -> 0.147; Electricity 0.147 -> 0.142; Traffic 0.377 -> 0.367 | Helped |
| Add dynamic weight to TVNet (ablation) | TVNet · Weather, Electricity, Traffic; prediction lengths 96, 192, 336, 720 | MSE: Weather 0.251 -> 0.147; Electricity 0.185 -> 0.142; Traffic 0.391 -> 0.367 | Helped |
| Increase embedding dimension from 32 to 64 (ablation) | TVNet · ETTh1, Weather, Electricity long-term forecasting | MSE: ETTh1 0.369 -> 0.371; Weather 0.149 -> 0.147; Electricity 0.139 -> 0.142 | Mixed |
| Increase 3D-block count from 2 to 3 (ablation) | TVNet · ETTh1, Weather, Electricity long-term forecasting | MSE: ETTh1 0.380 -> 0.371; Weather 0.153 -> 0.147; Electricity 0.146 -> 0.142 | Helped |
Surprising findings
- On several long-term forecasting datasets (Traffic, Exchange, Electricity in some rows) TVNet is not best, e.g. PatchTST and MICN/ModernTCN win on Traffic and Electricity MSE.
Limitations
- Several reported baseline comparisons show TVNet not best (e.g., Traffic, Electricity), and the paper claims SOTA more broadly.
- Theoretical proof that dynamic weights reduce error is a simple least-squares argument that assumes optimal per-patch alpha, not a generalization result.
- Anomaly detection and classification results are reported only as summary numbers in figures in the extracted text.
Experimental setup
Datasets: Weather Traffic Electricity Exchange ILI ETTh1 ETTh2 ETTm1 ETTm2 M4 UEA classification archive (10 subsets) SMD
Benchmarks: Weather Traffic Electricity Exchange ILI ETTh1 ETTh2 ETTm1 ETTm2 M4 UEA SMD
Metrics: MSE MAE SMAPE MASE OWA accuracy F1-score
Baselines: PatchTST iTransformer Crossformer FEDformer RLinear MTS-Mixer DLinear TimesNet MICN ModernTCN RMLP U-Mixer
Builds on: dynamic convolution TimesNet PatchTST
Closest papers by meaning
- MPTSNet: Integrating Multiscale Periodic Local Patterns and Global Dependencies for Multivariate Time Series Classification (AAAI 2025 · similarity 0.86)
- Time Series Representations with Hard-Coded Invariances (ICML 2025 · similarity 0.86)
- Fully-Connected Spatial-Temporal Graph for Multivariate Time-Series Data (AAAI 2024 · similarity 0.85)
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (NeurIPS 2025 · similarity 0.85)
- TSLANet: Rethinking Transformers for Time Series Representation Learning (ICML 2024 · similarity 0.85)
- Segment, Shuffle, and Stitch: A Simple Layer for Improving Time-Series Representations (NeurIPS 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.