Six-Week Weather Forecasts Improved Through Machine Learning-Based Framework Advancement
Researchers have developed a machine learning-powered probabilistic correction framework that significantly improves two- to six-week weather forecasts.
To improve the accuracy of weather forecasts predicting two to six weeks ahead, a new probabilistic correction framework has been introduced that calibrates the model using historical forecasts and observations.
The Importance of Six-Week Weather Forecasts
Six-week weather forecasts, covering two to six weeks in advance, are critical for agricultural planning, water resource management, disaster preparedness, and energy management.
However, achieving high accuracy during this timeframe is a very difficult process due to the accumulation of model errors, especially in precipitation forecasts.
How the Probabilistic Correction Framework Works
To overcome this predictability barrier, the probabilistic correction framework developed by researchers adaptively corrects the model's distribution using historical forecasts and weather observations.
The system incorporates two complementary machine learning models to adjust the location, date, and quartile values of forecasts by taking recent weather trends into account.
Models Providing Performance Improvement
Applied to industry-leading dynamic ensembles such as ECMWF, these models, named Debias++ and Persistence++, consistently improve low-scoring raw materials.
This developed method has tripled positive precipitation skill and transformed temperature and pressure forecasts with negative skill into successful predictions.
Integration with Artificial Intelligence Models
The probabilistic correction framework not only performs superiorly to standard operational correction protocols.
At the same time, it empowers state-of-the-art artificial intelligence six-week forecasting models like AIFS-SUBS within ECMWF, doubling the average skill across variables and lead times.