Bridging short- and medium-range weather forecasting with machine learning
Nested-EAGLE model integrates short- and medium-range forecasts, significantly reducing MSE over CONUS.
Key Findings
Methodology
Nested-EAGLE combines global and regional weather data, incorporating high-resolution regional analysis into training via nesting. It uses a graph transformer encoder and decoder, and a sliding-window transformer processor, focusing on high-resolution predictions over CONUS.
Key Results
- Nested-EAGLE significantly reduces MSE in near-surface and low-level forecasts over CONUS, improving prediction capability by 78 hours compared to GFS and HRRR.
- While less skillful in precipitation prediction than HRRR, Nested-EAGLE accurately locates storms at longer leads.
- The model remains competitive globally, with notable performance in near-surface variables over CONUS.
Significance
This study demonstrates the potential of integrating short- and medium-range forecasting systems through the Nested-EAGLE model, particularly for applications in CONUS. It offers a more unified approach to weather forecasting, reducing discrepancies between different systems and improving accuracy and utility.
Technical Contribution
Nested-EAGLE significantly enhances near-surface prediction accuracy by nesting high-resolution regional data. Unlike existing single-resolution models, it maintains efficient prediction capabilities globally and achieves significant skill improvements in CONUS.
Novelty
Nested-EAGLE is the first to embed high-resolution regional analysis data into a global weather model, significantly improving near-surface prediction accuracy, especially in CONUS.
Limitations
- The model's performance in precipitation prediction is less than HRRR, mainly due to blurring effects from deterministic training.
- It fails to propagate HRRR data skill improvements globally.
Future Work
Future work will extend skill improvements beyond CONUS and improve precipitation representation, potentially by introducing probabilistic training methods to enhance precipitation prediction accuracy.
AI Executive Summary
Weather forecasting has traditionally relied on separate short- and medium-range systems, each optimized for specific phenomena. However, this separation leads to complexity and inconsistency in predictions. The Nested-EAGLE model offers a more unified approach by integrating global and regional weather data, significantly reducing mean-squared error over CONUS while remaining competitive globally.
The Nested-EAGLE model employs a graph transformer encoder and decoder, and a sliding-window transformer processor, focusing on high-resolution predictions over CONUS. By nesting high-resolution regional analysis data, the model excels in near-surface and low-level variable predictions, particularly in wind speed and temperature.
Despite being less skillful in precipitation amounts than HRRR, Nested-EAGLE accurately locates storms at longer leads, showcasing its potential in weather forecasting. Future research will focus on extending skill improvements beyond CONUS and enhancing precipitation representation.
Deep Analysis
Background
Weather forecasting has long relied on independent short- and medium-range systems like NOAA's GFS and HRRR. These systems are optimized for specific phenomena, but their separation leads to complexity and inconsistency in predictions. Recent advances in machine learning offer new possibilities for integrating these systems.
Core Problem
Existing weather forecasting systems are separated into short- and medium-range predictions, leading to complexity and inconsistency. Effectively integrating these systems to improve accuracy and utility is a significant research challenge.
Innovation
The Nested-EAGLE model integrates short- and medium-range forecasting systems by nesting high-resolution regional analysis data. Unlike traditional single-resolution models, it maintains efficient prediction capabilities globally and achieves significant skill improvements in CONUS.
Methodology
- �� Uses graph transformer encoder and decoder to process global and regional weather data.
- �� Nests high-resolution regional analysis data to enhance near-surface prediction accuracy.
- �� Employs a sliding-window transformer processor focusing on high-resolution predictions over CONUS.
Experiments
Experiments used NOAA's HRRR and GFS data for training and testing. By comparing predictions from Nested-EAGLE, ML-GFS-Base, and traditional physical models, the model's capability in near-surface and low-level variable predictions was evaluated.
Results
Nested-EAGLE excels in near-surface predictions over CONUS, significantly reducing mean-squared error. While less skillful in precipitation amounts than HRRR, it accurately locates storms at longer leads.
Applications
The Nested-EAGLE model can be used to improve weather forecasting accuracy, particularly in CONUS. It offers a more unified approach to weather forecasting, reducing discrepancies between different systems.
Limitations & Outlook
The model's performance in precipitation prediction is less than HRRR, mainly due to blurring effects from deterministic training. Future work will focus on improving precipitation representation and extending skill improvements beyond CONUS.
Plain Language Accessible to non-experts
Imagine you're cooking in a kitchen. Traditional weather forecasting is like using different pots for different dishes, each with its own temperature and timing. The Nested-EAGLE model is like a smart cooker that can control multiple pots simultaneously, ensuring all dishes are perfectly cooked at the same time. By combining global and regional data, Nested-EAGLE acts like a smart cooker that automatically adjusts to the needs of each dish, ensuring optimal taste and texture.
ELI14 Explained like you're 14
Imagine you're playing a super complex weather simulation game. Traditional forecasting is like different levels in the game, each with its own rules and goals. The Nested-EAGLE model is like a super boss that can control multiple levels at once, making the whole game smoother and more fun. By combining global and regional data, Nested-EAGLE acts like a super boss that automatically adjusts to the needs of each level, ensuring the best gaming experience!
Glossary
Nested-EAGLE
A machine learning model combining global and regional weather data to improve forecasting accuracy.
Used in this paper to integrate short- and medium-range weather predictions.
HRRR
NOAA's regional weather forecasting system providing high-resolution short-term forecasts.
Used as regional data for training the Nested-EAGLE model.
GFS
NOAA's global weather forecasting system providing medium-range forecasts.
Used as global data for training the Nested-EAGLE model.
Graph Transformer
A deep learning model for processing graph-structured data.
Used as the encoder and decoder in the Nested-EAGLE model.
Mean-Squared Error
A loss function for evaluating model performance by calculating the square difference between predicted and true values.
Used to assess the prediction accuracy of the Nested-EAGLE model.
Open Questions Unanswered questions from this research
- 1 How to effectively propagate HRRR data skill improvements globally?
- 2 How to improve precipitation prediction accuracy, especially in extreme weather events?
Applications
Immediate Applications
Weather Forecast Improvement
The Nested-EAGLE model can be used to enhance weather forecasting accuracy in CONUS, aiding meteorologists in better predicting storms and precipitation events.
Long-term Vision
Global Weather Forecast Optimization
By improving the model's global applicability, Nested-EAGLE has the potential to become a standard tool for global weather forecasting, helping countries better address challenges posed by climate change.
Abstract
The National Oceanic and Atmospheric Administration (NOAA) employs independent prediction systems for distinct forecast products. While some separation is practical, we argue that combining short- and medium-range weather into a single prediction system would provide the public with a useful distillation of global weather and its impacts. To this end, we present Nested-EAGLE (Experimental Artificial intelligence Global and Limited-area Ensemble): a 0.25° global weather model with a 6 km refinement over the Contiguous United States (CONUS). The model achieves significantly lower mean-squared error in near-surface and low-level quantities over CONUS compared to NOAA's Global Forecast System and High-Resolution Rapid Refresh (HRRR), while remaining competitive throughout the rest of the global atmosphere. We show that the skill gains for near-surface fields stem from incorporating high-resolution regional analysis data into training through the nesting process. Forecasts of precipitation amounts are less skillful than those from HRRR, owing to deterministic training. However, we show that Nested-EAGLE provides the most accurate forecasts of storm locations at longer leads, despite blurred extrema. Our results motivate future work to extend the skill gains beyond CONUS and improve precipitation representation.