SDGF: Fusing Static and Multi-Scale Dynamic Correlations for Multivariate Time Series Forecasting
SDGF: Fusing Static and Multi-Scale Dynamic Correlations for Multivariate Time Series Forecasting
Authors: Shaoxun Wang, Xingjun Zhang, Qianyang Li, Jiawei Cao, Zhendong Tan Date: 2026-04-21 Paper ID: openalex:2509.18135
Summary
SDGF addresses the limitations of existing multivariate forecasting models in capturing complex, multi-scale inter-series dependencies. The architecture employs a dual-path approach: a static graph path based on prior knowledge for stable dependencies and a dynamic graph path driven by multi-level wavelet decomposition for multi-scale feature extraction. An attention-gated module integrates these sources, while a multi-kernel dilated convolutional network further refines temporal pattern modeling. Experiments across standard benchmarks show that this fusion strategy significantly improves forecasting accuracy.
Key Contributions
- Introduces SDGF, a dual-path graph fusion network that captures both long-term stable dependencies via static graphs and evolving associations across temporal scales.
- Proposes an attention-gated fusion module to intelligently integrate static prior knowledge with multi-scale dynamic graph features.
- Demonstrates superior performance over state-of-the-art baselines on multiple real-world multivariate time series forecasting benchmarks.
Open Questions & Future Work
Key Concepts
- static-dynamic-graph-fusion: A dual-path graph learning framework that combines static prior-based graphs with multi-scale, adaptively learned dynamic graphs for multivariate time series forecasting.
Archivist Review
I approved the ‘Static-Dynamic Graph Fusion’ concept because it defines a generalizable dual-path architectural pattern for balancing static priors and dynamic correlations in multivariate forecasting. I also approved a refined open question regarding the ‘Latent Inter-series Dependency Bottleneck’, focusing the research challenge on the limitations of current static/dynamic fusion paradigms when faced with non-stationary, latent dependencies. I rejected the initial open question to replace it with a more concise version suited for the vault.
Approved Concepts
- Static-Dynamic Graph Fusion: The core novelty of the paper is the specific dual-path graph structure learning that merges static prior knowledge with adaptively learned multi-scale dynamic dependencies.
Approved Open Questions
- Latent Inter-series Dependency Bottleneck: The effectiveness of multivariate forecasting is restricted by the rigidity of current graph structure learning methods, limiting the ability to adapt to complex, evolving, and latent inter-series associations.
Rejected Candidates
- [open_question] Capturing Complex Evolving Dependencies (
complex-evolving-inter-series-dependencies) - other: The suggested title and description were reframed into a more specific, technical open-question slug that better highlights the bottleneck nature of the problem.
Links
Metadata & Links
- url
- https://arxiv.org/abs/2509.18135
- paper_id
- 2509.18135
- paper_source
- openalex
- domain
- time-series
- tags
- time-series-forecastinggraph-neural-networksmultivariate-time-serieswavelet-analysisattention-mechanism
- architectures
-
- datasets
-
- concept_slugs
- static-dynamic-graph-fusion
- dataset_slugs
-
- skill
- TimeSeriesSkill
- created_at
- 2026-04-24T05:50:06Z