Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Adaptive Spatio-Temporal Graphs with Self-Supervised Pretraining for Multi-Horizon Weather Forecasting

Yao Liu

TL;DR

The paper tackles the challenge of robust, high-resolution multi-variable weather forecasting in urban and complex regions, where traditional numerical models are computationally expensive and data-driven methods struggle with generalization. It proposes a spatio-temporal self-supervised learning framework that integrates a graph neural network for spatial reasoning, self-supervised pretraining, and a horizon-aware adaptation mechanism to handle multiple forecast horizons. Key contributions include a total loss combining spatio-temporal prediction, contrastive, and consistency terms, a dynamic GNN with attention and adaptive adjacency, and extensive experiments on ERA5 and MERRA-2 showing substantial improvements over NWP baselines and existing DL approaches across 24–168 hours, including focusing on Beijing and Shanghai. The results demonstrate label-efficient learning, cross-dataset generalization, and scalability for data-driven weather forecasting in practical deployment.

Abstract

Accurate and robust weather forecasting remains a fundamental challenge due to the inherent spatio-temporal complexity of atmospheric systems. In this paper, we propose a novel self-supervised learning framework that leverages spatio-temporal structures to improve multi-variable weather prediction. The model integrates a graph neural network (GNN) for spatial reasoning, a self-supervised pretraining scheme for representation learning, and a spatio-temporal adaptation mechanism to enhance generalization across varying forecasting horizons. Extensive experiments on both ERA5 and MERRA-2 reanalysis datasets demonstrate that our approach achieves superior performance compared to traditional numerical weather prediction (NWP) models and recent deep learning methods. Quantitative evaluations and visual analyses in Beijing and Shanghai confirm the model's capability to capture fine-grained meteorological patterns. The proposed framework provides a scalable and label-efficient solution for future data-driven weather forecasting systems.

Adaptive Spatio-Temporal Graphs with Self-Supervised Pretraining for Multi-Horizon Weather Forecasting

TL;DR

The paper tackles the challenge of robust, high-resolution multi-variable weather forecasting in urban and complex regions, where traditional numerical models are computationally expensive and data-driven methods struggle with generalization. It proposes a spatio-temporal self-supervised learning framework that integrates a graph neural network for spatial reasoning, self-supervised pretraining, and a horizon-aware adaptation mechanism to handle multiple forecast horizons. Key contributions include a total loss combining spatio-temporal prediction, contrastive, and consistency terms, a dynamic GNN with attention and adaptive adjacency, and extensive experiments on ERA5 and MERRA-2 showing substantial improvements over NWP baselines and existing DL approaches across 24–168 hours, including focusing on Beijing and Shanghai. The results demonstrate label-efficient learning, cross-dataset generalization, and scalability for data-driven weather forecasting in practical deployment.

Abstract

Accurate and robust weather forecasting remains a fundamental challenge due to the inherent spatio-temporal complexity of atmospheric systems. In this paper, we propose a novel self-supervised learning framework that leverages spatio-temporal structures to improve multi-variable weather prediction. The model integrates a graph neural network (GNN) for spatial reasoning, a self-supervised pretraining scheme for representation learning, and a spatio-temporal adaptation mechanism to enhance generalization across varying forecasting horizons. Extensive experiments on both ERA5 and MERRA-2 reanalysis datasets demonstrate that our approach achieves superior performance compared to traditional numerical weather prediction (NWP) models and recent deep learning methods. Quantitative evaluations and visual analyses in Beijing and Shanghai confirm the model's capability to capture fine-grained meteorological patterns. The proposed framework provides a scalable and label-efficient solution for future data-driven weather forecasting systems.

Paper Structure

This paper contains 17 sections, 14 equations, 9 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Traditional Numerical and Statistical Methods
  4. Deep Learning for Weather Forecasting
  5. Methodology
  6. Model Overview
  7. Self-supervised Learning Framework
  8. Spatio-temporal Adaptation Mechanism
  9. Graph Neural Network (GNN) Module
  10. Experimental Results and Analysis
  11. Data and Research Areas
  12. Experiment Setup
  13. Compared Methods
  14. Experimental Results
  15. Visualization Analysis
  16. ...and 2 more sections

Figures (9)

  • Figure 1: Overview of the proposed spatio-temporal self-supervised learning model for robust weather forecasting
  • Figure 2: Model Robustness in Beijing Using MERRA-2 Reanalysis Data. The plot compares model prediction errors across temperature, wind speed, humidity, and pressure for Beijing. Narrower distributions indicate higher robustness
  • Figure 3: Model Robustness in Shanghai Using MERRA-2 Reanalysis Data. The plot compares model prediction errors across temperature, wind speed, humidity, and pressure for Shanghai. Narrower distributions indicate higher robustness
  • Figure 4: Model Robustness in Beijing Using ERA5 Reanalysis Data. The plot compares model prediction errors across temperature, wind speed, humidity, and pressure for Beijing. Narrower distributions indicate higher robustness
  • Figure 5: Model Robustness in Shanghai Using ERA5 Reanalysis Data. The plot compares model prediction errors across temperature, wind speed, humidity, and pressure for Shanghai. Narrower distributions indicate higher robustness
  • ...and 4 more figures