What is it about?
Long-range weather forecasting, particularly on the subseasonal to seasonal (S2S) scale—weeks to months in advance—is notoriously difficult. Weather models, like NOAA’s Global Ensemble Forecast System (GEFS), often make significant errors, especially when predicting complex tropical systems like the Madden-Julian Oscillation (MJO) or the Boreal Summer Intraseasonal Oscillation (BSISO). These systems can influence weather in distant regions, like the United States, through atmospheric ripple effects known as teleconnections. Instead of simply predicting the weather, this study focuses on predicting where and when the GEFS model is likely to make errors. By identifying these “errors of opportunity,” researchers use neural networks—advanced machine learning models—to uncover the atmospheric conditions that lead to these errors. For instance, certain phases of the MJO and BSISO are associated with predictable patterns of model inaccuracy. The study reveals specific examples, such as the GEFS model overestimating high-altitude atmospheric pressures in the Pacific Northwest during spring and underestimating them in Northwest Mexico during summer. It also highlights errors in summer precipitation forecasts in the Midwest, particularly tied to BSISO phases 1 and 2. By understanding why these errors happen, the researchers show how machine learning tools can improve forecast accuracy. Ultimately, this work sheds light on how advanced neural networks and error analysis can help refine long-range weather forecasts, making them more reliable for applications like disaster preparedness, agriculture, and energy planning.
Featured Image
Photo by NOAA on Unsplash
Why is it important?
Weather forecasts are an essential tool for people, businesses, and governments worldwide. For example, farmers need forecasts to decide when to plant crops, energy companies rely on them to manage resources, and communities use them to prepare for extreme weather events like hurricanes or heatwaves. While short-term forecasts—up to a week ahead—are fairly accurate, predicting the weather weeks to months in advance (subseasonal to seasonal or S2S) is much harder. Yet, these longer-term predictions are critical for planning ahead and avoiding costly surprises. This research is important because it focuses on improving the accuracy of S2S forecasts. Traditional weather models, like NOAA’s Global Ensemble Forecast System (GEFS), often make mistakes when predicting complex weather patterns. These errors can snowball, leading to inaccurate predictions for important variables like rainfall, temperature, and atmospheric pressure. For instance, mistakes in forecasting tropical systems, such as the Madden-Julian Oscillation (MJO), can affect weather across the U.S. through atmospheric ripple effects called teleconnections. By using advanced neural networks, the study doesn’t just identify where the model gets the weather wrong; it also explains why these mistakes happen. Knowing the "why" can help scientists improve weather models in the future, leading to better long-term forecasts. This matters because more reliable S2S forecasts can help farmers reduce crop loss, utilities better manage energy grids, and governments better prepare for extreme weather events. This work is a step toward making forecasting tools smarter and more dependable, benefiting society as a whole.
Perspectives
This research goes beyond traditional forecasting by focusing on the root cause of errors, which is often overlooked. Rather than simply improving predictions of weather outcomes, it targets the accuracy of the models themselves, creating an entirely new pathway for progress in long-range forecasting. This is a big deal because weather models are only as good as their ability to represent real-world dynamics. By zeroing in on when and why the GEFS model struggles, the study adds a new layer of understanding that bridges the gap between current limitations and future possibilities. What makes this particularly exciting is the use of neural networks not just as predictive tools but as analytical ones. Explainability in machine learning is notoriously tricky, and applying it to atmospheric science opens up a world of potential. For example, knowing that certain phases of the BSISO or MJO make the model prone to errors allows researchers to adjust inputs, recalibrate processes, or even design future models that "learn" from these patterns. In practical terms, this approach is a game changer for industries that rely on long-term forecasts. Better predictions of where models fail can lead to smarter contingency planning. For example, energy companies could adapt their strategies during high-risk periods, and emergency responders could better anticipate resource needs before disasters strike. By shifting the focus from just predicting weather to understanding and addressing model weaknesses, this research offers a smarter, more proactive way to tackle forecasting challenges.
Jack Cahill
Colorado State University
Read the Original
This page is a summary of: Errors of Opportunity: Using Neural Networks to Predict Errors in the Global Ensemble Forecast System (GEFS) on S2S Time Scales, Weather and Forecasting, December 2024, American Meteorological Society,
DOI: 10.1175/waf-d-23-0125.1.
You can read the full text:
Contributors
The following have contributed to this page
