The National Oceanic and Atmospheric Administration (NOAA) of the United States has recently officially launched a new generation of global weather forecast models based on artificial intelligence, claiming to achieve faster and more accurate weather forecasts while significantly reducing computing power consumption. This marks a major leap in the US weather forecast system from traditional physical models to data-driven AI systems. NOAA said the technology was put into operation early Wednesday morning and was an important step in its efforts to modernize the U.S. weather prediction system.
These AI models were developed and deployed by NOAA’s Center for Environmental Forecasts in collaboration with the National Weather Service and are positioned to supplement, not replace, existing numerical forecast models. National Weather Service spokesperson Erica Grow Cei told the media that part of the training data for the currently running machine learning models comes from traditional numerical models that are still in use, and those models that rely on complex physical equations are still one of the important information sources for AI training.
For a long time, NOAA's core forecast tool has been the Global Forecast System (GFS). This basic physical model simulates atmospheric behavior through mathematical equations and generates data on multiple elements such as temperature, wind speed, precipitation, ozone, and soil moisture. It is composed of multiple subsystems such as the land surface, ocean, and atmosphere. It collaborates to form a whole. In order to mitigate the systematic bias of the GFS, NOAA also previously built the "Global Ensemble Forecast System" (GEFS) to cover the uncertainty of different weather scenarios through multiple simulations.
Daryl Kleist, deputy director of NOAA's Environmental Forecast Center, said that the new generation of AI models are trained based on the data accumulated by these traditional models over the years. He pointed out that the significant improvement in forecasting skills of these AI models is largely due to the "analytical field" data used in their training, and these analytical data are mainly derived from the old numerical model framework.
In terms of computing power requirements, NOAA estimates that the new AI system can reduce computing resource usage by 91% to 99% compared with traditional forecast models, significantly reducing the dependence of real-time business forecasts on supercomputing clusters. At the same time, these models are expected to extend the effective forecast time by 18 to 24 hours while maintaining or improving accuracy. Kleist also reminded that the energy consumption calculated here is the energy consumption during the running stage of the model, and does not include the large energy investment required for the early AI training itself.
The AI forecasting system launched this time consists of three core models. The first is the Artificial Intelligence Global Forecast System (AIGFS), which officials describe as a new global model that uses AI technology to generate weather forecasts in a faster and more efficient way. According to data given by NOAA, AIGFS only requires about 0.3% of the computing resources of traditional GFS to complete a 16-day global forecast, and the running time is about 40 minutes, which means that operational forecasters can obtain updated numerical guidance earlier.
The second model is the "Artificial Intelligence Global Ensemble Forecast System" (AIGEFS), which introduces ensemble ideas on the basis of AIGFS. It no longer only gives a single deterministic result, but generates a series of possible evolution paths to quantify the uncertainty in weather forecasts. The third model "Hybrid-GEFS" integrates new AI technology with NOAA's existing GEFS ensemble system, aiming to further use AI to optimize the representation of uncertainty and forecast accuracy while retaining the advantages of the traditional ensemble system.
NOAA emphasized that this series of AI models are still in the continuous iteration stage, and the scientific research team is focusing on improving its performance in high-impact weather forecasts such as hurricanes, and further improving the range of possible scenarios given by the ensemble system. The agency believes that as these models continue to improve, AI is expected to play an increasingly critical supporting role in future extreme weather warnings and medium- and long-term forecasts.