SHANGHAI — Envision Group says its "Tianji" weather model reached the first tier of the international WeatherBench 2 benchmark and took first place in several domestic evaluations, results the company describes as a domestic and international "double championship."
Envision describes Tianji as the world's first physical-AI weather model, developed to compete with comparable systems from global technology companies including Google, Microsoft and Huawei. The model was presented at the "Energy Storage AI Innovation Forum" held during the 2026 Inclusion·Bund Summit in Shanghai on Sept 9, according to Envision and Chinese reports by Beijing Daily's Jingbao website and Science and Technology Daily.
The performance claims come from the company and the Chinese outlets that covered the forum. The reports do not include the benchmark's detailed methodology or any independent replication of the results.
Global benchmark
In the WeatherBench 2 evaluation, only Tianji and Google's WeatherNext 2 model entered the benchmark's first tier, according to the reports. Tianji was said to lead Google on wind speed and precipitation, indicators the reports describe as more important to the energy industry.
Domestic evaluations
Envision says Tianji took first place overall in an AI weather-forecast model demonstration program organized by China's National Meteorological Center, and first place in total score in a State Grid AI load-forecasting competition. The State Grid contest covered load forecasting for the Central China power grid, four provincial grids in the region and provincial total-load forecasting across State Grid areas, the reports said.
Previously, the model topped a China Southern Power Grid AI load-forecasting technology competition with 132.34 points, leading key provinces including Guangdong by more than 20 points, according to the reports. In Central China — Hubei, Jiangxi, Henan and Hunan — Tianji was reported to hold first place in forecasting, with accuracy above 98%; that figure appears in one of the reports.
Stated capability and typhoon tests
Envision says the model can forecast, at minute-level intervals, conditions for any region worldwide up to 45 days ahead, at a resolution of 5 kilometers and 10 minutes.
During this year's typhoon season, the reports say, Tianji was used through three typhoons named in the Chinese-language accounts as Hongxia, Baihaitun and Shade'er. It predicted typhoon paths and impact windows 72 hours in advance and provided minute-level wind forecasts as the storms passed, giving operations teams staged instructions before, during and after each event — including when to stop turbines and when to restart them. Envision says the capability has been tested at onshore and offshore wind farms.
Energy applications
Zheng Ying, who heads the Tianji weather model, said the system draws on global meteorological data, satellite remote sensing and first-hand measurements from new-energy sites, parks and factories, allowing it to capture changes in temperature, humidity, rainfall and wind speed and to combine them with the operating characteristics of different regions and energy scenarios.
In storage, Envision says its AI energy-storage system, built on the Tianji model, held first place in trading-forecast accuracy for consecutive months at multiple sites in Inner Mongolia. Verified across benchmark projects in Shandong, Inner Mongolia and Jiangsu, the system delivered a 20% increase in comprehensive station revenue and an 8% improvement in full-lifecycle internal rate of return, according to the company.
The model's training relies on the AI computing center of Envision's Mission Gobi program, while its weather forecasts feed back into variable wind and solar generation — a loop the company calls "Energy for AI, AI for Energy."
A separate report by Securities Daily described the model's use at Envision's Chifeng industrial park, where, supported by Envision's AI power system, power generation and hydrogen production respond in seconds: Tianji predicts changes in wind and solar output and automatically raises electrolyzer load before wind speeds increase.
What is confirmed, what is not
The forum's date and location, the model's launch there and the company's stated rankings all come from Envision and the Chinese outlets that covered the event. The reports do not include the benchmark's published results and methodology, comment from the evaluation organizers, or independent verification of the 98%-plus accuracy figure and the domestic competition scores. No timeline has been announced for further third-party evaluation or for publication of the underlying data.