Adaptive spatio-temporal graphs with self-supervised pretraining for multi-horizon weather forecasting
Adaptive spatio-temporal graphs with self-supervised pretraining for multi-horizon weather forecasting
Authors: Y. A. Liu Date: 2026-04-20 Paper ID: openalex:2511.00049
Summary
This paper addresses the complexity of multi-horizon weather forecasting by proposing a self-supervised learning framework that incorporates graph neural networks for spatial reasoning. By integrating a spatio-temporal adaptation mechanism, the model improves generalization across various time scales and meteorological variables. Empirical evaluations on the ERA5 and MERRA-2 datasets demonstrate superior performance over traditional numerical weather prediction and existing deep learning benchmarks. The approach provides a scalable and label-efficient solution for fine-grained atmospheric pattern prediction.
Key Contributions
- Introduces a self-supervised pretraining framework for learning spatio-temporal representations in weather data.
- Develops a spatio-temporal adaptation mechanism that enhances predictive accuracy across multiple forecasting horizons.
- Outperforms traditional numerical weather prediction and current deep learning baselines on ERA5 and MERRA-2 benchmarks.
Open Questions & Future Work
Archivist Review
The paper provides solid benchmark results on standard climate datasets but lacks novel, reusable architectural innovations; therefore, no new concepts were approved. ERA5 and MERRA-2 were approved as they are foundational, widely used datasets in this domain. The open question regarding extreme weather performance was approved as it identifies a critical, distinct bottleneck in data-driven forecasting that persists across various architectures.
Approved Open Questions
- Extreme Weather Event Forecasting: Improving performance on extreme weather events is vital for the societal utility of forecasting systems, specifically for disaster preparedness and climate risk management.
Rejected Candidates
- [open_question] Dynamic Graph Construction Methods (
dynamic-graph-construction-weather) - generic: The question of dynamic vs. static graph construction is a well-known, generic architectural challenge in GNN-based spatiotemporal forecasting rather than a specific research gap highlighted as a priority by the paper.
Datasets
Links
Metadata & Links
- url
- https://arxiv.org/abs/2511.00049
- paper_id
- 2511.00049
- paper_source
- openalex
- domain
- time-series
- tags
-
- architectures
-
- datasets
- era5merra-2
- concept_slugs
-
- dataset_slugs
- era5merra-2
- skill
- TimeSeriesSkill
- created_at
- 2026-04-23T05:47:40Z