Envision said on 30 October that its Tianji weather foundation model ranked first in the latest AI Weather Quest, run by the European Centre for Medium-Range Weather Forecasts (ECMWF). In an announcement issued in Shanghai, the green technology company said the result covered forecasts for days 19 to 25 and put Tianji ahead of nearly 20 teams. The competition checks predictions as later weather becomes available, so the ranking concerns forecasts made in an ongoing test rather than a single demonstration.
Key points
- Envision says Tianji led the assessed temperature, pressure and precipitation measures for days 19 to 25.
- The model entered as Arcana in a contest that verifies forecasts against subsequent weather.
- NAI 500 reports a different lead window, days 26 to 33.
ECMWF checks forecasts against subsequent weather
Teams submit forecasts each week, and organisers compare them with the weather that follows before updating a public leaderboard, according to Envision. That continuing process exposes entries to changing conditions: a prediction is recorded before the conditions it describes can be observed. Envision says Tianji competed under the name Arcana and placed first overall in the edition it announced.
For days 19 to 25, Envision says Tianji led every assessed measure: near-surface temperature averaged across variables, mean sea level pressure and precipitation. The contest examines weekly mean temperature and pressure and weekly accumulated rain, using probabilistic forecasts that account for uncertainty rather than giving only one outcome. Temperature and precipitation are scored over grid cells predominantly covered by land; pressure is assessed globally. Those choices matter because the measures describe conditions across a week and, depending on the variable, across different geographic areas.
Planning an outdoor event weeks ahead could use an early estimate of how wet the wider week might be, rather than a prediction of rainfall at the event’s particular hour.
The 19-to-25-day window sits beyond the 15-day medium-range horizon described in Envision’s announcement. At that distance, a weekly probability is a different kind of answer from a detailed forecast for a particular afternoon: it deals with the broad conditions a week might bring. Microsoft researchers published a method to improve forecasts two to six weeks ahead, AI Affairs previously reported, addressing another part of this longer-range forecasting problem.
Zheng describes Tianji’s data network
Robert Zheng, head of Envision’s Tianji Weather Foundation Model, describes subseasonal forecasting as looking between two weeks and two months ahead. He says the growing uncertainty over that period places greater demands on a model’s ability to identify developing weather patterns. Envision calls Tianji a physics-AI model and says it draws on petabyte-scale data with high resolution in both space and time, alongside a three-dimensional “space–air–ground” data network.
That description points to the material feeding Tianji, rather than to a measured advantage from any individual part of its design. Weather changes across both a map and a calendar: observations taken at different locations and times give a forecasting system the history from which it can learn. Envision also says high-performance computing supports Tianji’s short-term and global forecasts, and claims improved accuracy for wind speed and irradiance in key regions.
The range of variables matters to Envision’s energy applications. Wind speed and irradiance concern conditions relevant to wind and solar generation, while the AI Weather Quest result covers temperature, sea level pressure and rain several weeks ahead. Envision says Tianji spans predictions made minutes ahead through to subseasonal assessments, though its reported contest placing is for a particular forecast window and set of measures.
NAI 500 gives a different forecast window
NAI 500 reported on 30 October that Tianji led the three assessed measures for days 26 to 33. Envision’s announcement puts its first-place results for temperature, pressure and precipitation in the 19-to-25-day window. NAI 500 also identifies Arcana as Tianji’s competition name.
Envision says Tianji earlier earned a top-tier placement in the WeatherBench 2 evaluation, on par with Google’s WeatherNext2 model. The company also says Tianji placed first in the deterministic forecast track of an AI weather-model demonstration programme organised by China’s National Meteorological Center. A deterministic forecast gives a single predicted outcome, whereas the AI Weather Quest measures described by Envision concern probabilistic weekly forecasts. For developers building around weather predictions, the distinction affects the kind of answer a model supplies, even when both are described as forecasts.
Envision says Tianji also placed first in State Grid’s AI load forecasting competition and in a business competition run by China Southern Power Grid.
