Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FengWu-GHR: Learning the Kilometer-scale Medium-range Global Weather Forecasting

Tao Han, Song Guo, Fenghua Ling, Kang Chen, Junchao Gong, Jingjia Luo, Junxia Gu, Kan Dai, Wanli Ouyang, Lei Bai

TL;DR

FengWu-GHR tackles the challenge of kilometer-scale global weather forecasting with a data-driven model trained on LR reanalysis and deployed at 0.09° resolution. It introduces a transformer-based meta model calibrated via Spatial Identical Mapping Extrapolate (SIME) to upscale inputs, augmented by Decompositional and Combinational Transfer Learning (DCTL) with Regional Enhanced Simulation modules and per-step Low-Rank Adaptation (LoRA) to enable long-horizon forecasts. Hindcasts for 2022 show FengWu-GHR surpassing IFS-HRES, and extensive evaluations on stations and extreme events demonstrate improved accuracy, reduced bias drift, and enhanced reliability at high resolution. The approach offers a plug-and-play path to high-resolution ML-based NWP, leveraging LR priors and targeted HR adaptations to overcome data and computational constraints.

Abstract

Kilometer-scale modeling of global atmosphere dynamics enables fine-grained weather forecasting and decreases the risk of disastrous weather and climate activity. Therefore, building a kilometer-scale global forecast model is a persistent pursuit in the meteorology domain. Active international efforts have been made in past decades to improve the spatial resolution of numerical weather models. Nonetheless, developing the higher resolution numerical model remains a long-standing challenge due to the substantial consumption of computational resources. Recent advances in data-driven global weather forecasting models utilize reanalysis data for model training and have demonstrated comparable or even higher forecasting skills than numerical models. However, they are all limited by the resolution of reanalysis data and incapable of generating higher-resolution forecasts. This work presents FengWu-GHR, the first data-driven global weather forecasting model running at the 0.09$^{\circ}$ horizontal resolution. FengWu-GHR introduces a novel approach that opens the door for operating ML-based high-resolution forecasts by inheriting prior knowledge from a pretrained low-resolution model. The hindcast of weather prediction in 2022 indicates that FengWu-GHR is superior to the IFS-HRES. Furthermore, evaluations on station observations and case studies of extreme events support the competitive operational forecasting skill of FengWu-GHR at the high resolution.

FengWu-GHR: Learning the Kilometer-scale Medium-range Global Weather Forecasting

TL;DR

FengWu-GHR tackles the challenge of kilometer-scale global weather forecasting with a data-driven model trained on LR reanalysis and deployed at 0.09° resolution. It introduces a transformer-based meta model calibrated via Spatial Identical Mapping Extrapolate (SIME) to upscale inputs, augmented by Decompositional and Combinational Transfer Learning (DCTL) with Regional Enhanced Simulation modules and per-step Low-Rank Adaptation (LoRA) to enable long-horizon forecasts. Hindcasts for 2022 show FengWu-GHR surpassing IFS-HRES, and extensive evaluations on stations and extreme events demonstrate improved accuracy, reduced bias drift, and enhanced reliability at high resolution. The approach offers a plug-and-play path to high-resolution ML-based NWP, leveraging LR priors and targeted HR adaptations to overcome data and computational constraints.

Abstract

Kilometer-scale modeling of global atmosphere dynamics enables fine-grained weather forecasting and decreases the risk of disastrous weather and climate activity. Therefore, building a kilometer-scale global forecast model is a persistent pursuit in the meteorology domain. Active international efforts have been made in past decades to improve the spatial resolution of numerical weather models. Nonetheless, developing the higher resolution numerical model remains a long-standing challenge due to the substantial consumption of computational resources. Recent advances in data-driven global weather forecasting models utilize reanalysis data for model training and have demonstrated comparable or even higher forecasting skills than numerical models. However, they are all limited by the resolution of reanalysis data and incapable of generating higher-resolution forecasts. This work presents FengWu-GHR, the first data-driven global weather forecasting model running at the 0.09 horizontal resolution. FengWu-GHR introduces a novel approach that opens the door for operating ML-based high-resolution forecasts by inheriting prior knowledge from a pretrained low-resolution model. The hindcast of weather prediction in 2022 indicates that FengWu-GHR is superior to the IFS-HRES. Furthermore, evaluations on station observations and case studies of extreme events support the competitive operational forecasting skill of FengWu-GHR at the high resolution.
Paper Structure (13 sections, 6 equations, 5 figures)

This paper contains 13 sections, 6 equations, 5 figures.

Table of Contents

  1. Introduction
  2. FengWu-GHR
  3. Meta Model
  4. Spatial Identical Mapping Extrapolate
  5. Decompositional and Combinational Transfer Learning
  6. Low-rank Adaptation for Long-lead Forecasting
  7. Result
  8. Dataset
  9. Verificatoin Metrics
  10. Global weather forecast verification on operational analysis
  11. Overall forecast evaluation on station observation
  12. Extreme heat wave and winter storm
  13. Conclusions

Figures (5)

  • Figure 1: Structutr of FengWu-GHR. a) The first stage is to learn the physical laws from the long-term historical reanalysis data at a low resolution. b) An extrapolation method that enables pretrained LR models to operate on high-resolution analysis fields. c) A transfer learning algorithm that can complement the pretrained model to capture the small-scale weather phenomena. d) A low-rank adaptation that is implemented on the parameter level for each step to improve the forecast skill during long roll-out.
  • Figure 2: The predictive skills of FengWu-GHR and IFS-HRES in weather forecast during 2022. RMSE: the lower the better. ACC: the higher the better, Bias: the closer to zero the better. Activity: the lower the forecast activity the smoother the forecast. The forecast evaluation metric of FengWu-GHR and IFS-HRES are both computed against the operational analysis data at the grid of $0.09^\circ \times 0.09^\circ$. The x-axis in each sub-figure represents lead time, at a 6-hour interval over a 10-day forecast.
  • Figure 3: Comparison of RMSE for surface temperature prediction relative to station observations during 2022. The plot displays the performance of IFS-HRES (blue), Pangu-weather (cyan), and FwngWu-GHR (red).
  • Figure 4: Surface temperature forecast of (a) Pangu-weather with $0.25^{\circ}$ spatial resolution, (b) IFS-HRES, (c) FengWu-GHR. They are initialized at 2022-07-07T12:00:00 UTC. (d) Operational analysis: it could be regarded as the ground truth assimilated from the observation data at the valid time.
  • Figure 5: A severe winter storm event prediction in Dec. 2022. The temperature prediction at 2-meter at the New York surface weather station as a function of the forecast lead time in the Pangu-weather (a), IFS-HRES (b), and FengWu-GHR (c).