Spatial-Temporal Fusion Graph Neural Networks with Mixed Adjacency For Weather Forecasting

Rapidly accumulating, large-scale and long-term meteorological data provide unprecedented opportunities for data-driven meteorological models and fine-grained numerical weather prediction. Many existing approaches based on deep learning models, e.g., recurrent neutral networks and graph neural netwo...

Full description

Saved in:
Bibliographic Details
Published in:IEEE access 2025-01, Vol.13, p.1-1
Main Authors: Guo, Ang, Liu, Yanghe, Shao, Shiyu, Shi, Xiaowei, Feng, Zhenni
Format: Article
Language:English
Subjects:
Atmospheric modeling
Correlation
Feature extraction
Graph neural networks
Meteorological data
Meteorology
Monitoring
Numerical models
Numerical weather forecasting
Predictive models
Spatial-temporal fusion graph
Spatial-temporal time series prediction
Spatiotemporal data
Transformers
Weather forecasting
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Rapidly accumulating, large-scale and long-term meteorological data provide unprecedented opportunities for data-driven meteorological models and fine-grained numerical weather prediction. Many existing approaches based on deep learning models, e.g., recurrent neutral networks and graph neural networks, have been proposed for weather forecasting. However, the subtle spatial correlations hidden in the vast amount of historical meteorological data have not been fully explored, such as dynamic spatial correlation. In this paper, we propose STGAMAM, which integrates Spatial-Temporal fusion Graph neural networks with a novel Adjacency Matrix and self-Attention Mechanisms to capture both long-term temporal periodicity and short-term spatial-temporal dependencies based on mixed adjacency via graph attention networks and then makes fine-grained prediction on concatenated features which combines diverse correlations. Our approach is validated by extensive experiment on two real-world datasets, which demonstrates the superiority of the proposed method over existing methods.
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2025.3532473