Extending weather forecasts from hours to weeks with teleconnection-aware graph neural networks

Abstract Subseasonal forecasting is essential for agricultural planning, water-resource management, and early warning of natural hazards, yet remains challenging because forecast errors grow rapidly with lead time. Recent machine-learning models have achieved remarkable foresting skills, but training such models requires substantial computational resources and associated energy consumption. Here, we propose a parameter-efficient framework that reuses a pretrained medium-range weather model and extends its forecast horizon from 24 hours to four weeks. The architecture retains the pretrained spatial-attention modules while introducing lightweight temporal attention to model dependencies across multiple atmospheric states. To account for the influence of large-scale boundary conditions at subseasonal timescales, we further incorporate climate teleconnection indices through a difference-based graph module. This physically motivated component learns how large-scale teleconnection patterns modulate the temporal evolution of geographically remote areas. Experiments show that the proposed framework reduces root-mean-square error for most evaluated atmospheric variables from the two-week forecast horizon onward. The framework requires updating only 10.9% of the total model parameters while enabling the processing of substantially longer temporal sequences. These results further indicate the transferability of spatial representations learned from medium-range forecasting to extended-range prediction.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-13
DOI
https://doi.org/10.21203/rs.3.rs-11430274/v1
Primary Topic
Meteorological Phenomena and Simulations
Type
preprint

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Extending weather forecasts from hours to weeks with teleconnection-aware graph neural networks

Xiao Xiang Zhu, Shan Zhao, Zhitiong Xiong
1 links
Research Square
Meteorological Phenomena and Simulations
preprint

Extending weather forecasts from hours to weeks with teleconnection-aware graph neural networks

Xiao Xiang Zhu, Shan Zhao, Zhitiong Xiong
preprint en

Abstract

Abstract Subseasonal forecasting is essential for agricultural planning, water-resource management, and early warning of natural hazards, yet remains challenging because forecast errors grow rapidly with lead time. Recent machine-learning models have achieved remarkable foresting skills, but training such models requires substantial computational resources and associated energy consumption. Here, we propose a parameter-efficient framework that reuses a pretrained medium-range weather model and extends its forecast horizon from 24 hours to four weeks. The architecture retains the pretrained spatial-attention modules while introducing lightweight temporal attention to model dependencies across multiple atmospheric states. To account for the influence of large-scale boundary conditions at subseasonal timescales, we further incorporate climate teleconnection indices through a difference-based graph module. This physically motivated component learns how large-scale teleconnection patterns modulate the temporal evolution of geographically remote areas. Experiments show that the proposed framework reduces root-mean-square error for most evaluated atmospheric variables from the two-week forecast horizon onward. The framework requires updating only 10.9% of the total model parameters while enabling the processing of substantially longer temporal sequences. These results further indicate the transferability of spatial representations learned from medium-range forecasting to extended-range prediction.

Research Square
Technical University of Munich (DE)
Bundesministerium für Wirtschaft und Klimaschutz, Australian Government, European Commission, Bundesministerium für Umwelt, Naturschutz, nukleare Sicherheit und Verbraucherschutz
Climate action
Meteorological Phenomena and Simulations
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon