China’s national weather agency is about to put an AI-built ensemble forecast into daily use. The China Meteorological Administration (CMA) says its Fengqing Ensemble V1.0 system has passed its operational review and will “officially commence operations on November 10”, according to the CMA’s English announcement. The agency first announced the decision in Chinese on 26 October.
An ensemble forecast runs many slightly different versions of a forecast instead of one. Where the versions agree, forecasters can be more confident. Where they spread apart, the weather is harder to call. Doing this with traditional physics models takes a great deal of computing power, which is one reason weather centres are testing AI models that can produce many runs quickly.
In brief
- The China Meteorological Administration says its AI ensemble forecast system, Fengqing Ensemble V1.0, has passed review and starts operations on 10 November.
- Twice a day it will produce 51 global forecasts out to 360 hours, or 15 days, and the CMA says its key forecasts stay useful for 11 days.
- The CMA also launched two AI models for its Fengyun weather satellites, one to read typhoon images and one to warn of severe storms up to four hours ahead.
What Fengqing Ensemble produces
The system was built by the CMA’s National Meteorological Centre with the Xiong’an Institute of Meteorological Artificial Intelligence and the National Meteorological Information Centre. It was developed under the CMA’s Earth system forecasting strategy for 2025 to 2035, the Chinese announcement says.
Twice a day, it will produce 51 forecasts, called ensemble members, covering the whole globe out to 360 hours, or 15 days. Output comes in 6-hour steps on a grid of 0.25 degrees, roughly 28 kilometres at the equator. It covers wind, temperature, humidity and the height of pressure surfaces at 13 levels of the atmosphere, plus surface measures such as 10-metre wind, 2-metre temperature and sea-level pressure.
The CMA says the system “fills a gap” in its own work on AI ensemble forecasting. Until now, the agency had not developed an AI system of this kind in-house.
How the CMA says it performs
The CMA tested the system on independent samples from January 2025 to August 2026. It says the effective forecast lead time for global 500 hPa geopotential height reaches 11 days. That measure tracks the large pressure patterns in the middle of the atmosphere that steer weather systems, and is a standard yardstick for medium-range forecasts.
The agency adds that two common ensemble checks perform well. One, the Continuous Ranked Probability Score, measures how close the spread of forecasts sits to what actually happened. The other, the spread-error ratio, tests whether the ensemble’s uncertainty matches its real errors. The CMA has not published the scores themselves, nor comparisons with other centres’ systems. The Chinese announcement says the units involved will track how the products are used and keep improving the system.
AI models for Fengyun satellites
On 29 October, at the 2026 Fengyun satellite user conference, the CMA also released its first AI models for extreme weather built on Fengyun satellite data, according to the CMA.
The first reads typhoon cloud images. It combines several satellite channels to work out a storm’s position, structure, intensity and trend, and writes a text summary. It covers seven types of typhoon feature, uses more than 20 live satellite products, runs every 15 minutes and produces its text in 27 seconds. The CMA says it was used on this year’s typhoon No. 26.
The second aims at severe convective storms, the sudden thunderstorms behind flash floods and hail. It takes the past two hours of Fengyun-4 observations and predicts how the clouds will change over the next four hours, splitting that into the movement of large systems and the growth and decay of small ones. The CMA says this brings warning time forward to four hours. A research team of 49 researchers from 12 institutions, including Harbin Institute of Technology (Shenzhen), will work on further AI uses for the satellites.
Faster runs mean more members
This is TSN’s analysis. The main benefit of an AI ensemble is scale. Fifty-one global runs out to 15 days, twice a day, is a heavy load for a physics model. An AI model can produce them far faster once trained, which leaves room for more members or more frequent updates. That matters most for rare, high-impact events, where a forecaster needs to know how likely a storm is, not just whether one model shows it. The Fengyun models push the same idea into the short range, much as Vaisala’s NVIDIA-based lightning forecasts do, and they sit alongside other hazard work such as Mila’s wildfire mapping AI.
Self-reported skill, so far
The evidence so far comes from the CMA itself. The 11-day figure is from its own tests, and it has not released the scores or a comparison with existing ensembles, so it is hard to judge how Fengqing stacks up. AI weather models also learn from past data, which can make them weaker on extremes they rarely saw in training. Real use from 10 November will be the real test. As MIT and Microsoft’s longer-range work shows, machine learning can improve forecasts, but each system has to prove itself in daily service. For more on AI used for public good, see TSN’s Tech for Good hub.
Sources
- China Meteorological Administration, “CMA’s AI Global Short- to Medium-Range Ensemble Forecasting System ‘Fengqing Ensemble V1.0’ to Commence Operations”, 29 October 2026
- China Meteorological Administration, “中国气象局人工智能全球短中期集合预报系统—风清集合V1.0将业务运行” (in Chinese), 26 October 2026
- China Meteorological Administration, “风云气象卫星极端天气AI模型发布” (Fengyun satellite extreme-weather AI models released, in Chinese), 29 October 2026
Related stories
- Vaisala Says NVIDIA AI Stretches Lightning Warnings to Three Hours
- Mila and NRCan AI Maps Wildfire Hazard Hundreds of Times Faster
- MIT and Microsoft Researchers Use Machine Learning to Improve Weather Forecasts Two to Six Weeks Out
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
