Page 192 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

Use Case 2: AI-Powered High-Precision Weather Forecasting

Model

Organization: Alibaba Damo (Hangzhou) Technology Co., Ltd.

Country: China
Contact Person(s):

Yi Chen, elaine.cy@ alibaba -inc .com
Ting Wang, wt242148@ alibaba -inc .com

1 Use Case Summary Table

Item Details
Category Climate Change/Natural Disaster

Problem Addressed Although widely-recognized as the most effective weather forecast-
ing methods, Numerical Weather Prediction (NWP) is computationally
intensive and has limitations in forecasting timeliness and accuracy over
extended forecast periods. Based on numerous weather data, AI based
technology is expected to capture complex weather patterns and provide
economic and precise weather forecast.
Key Aspects of Solu- Established a high-precision regional forecasting model based on a
tion self-developed AI-driven global weather forecasting model to deliver
accurate weather forecasts with a 1-kilometer grid resolution on an hourly
basis, thereby meeting diverse predictive demands from various indus-
tries. [1]

Technology MAE, Swin Transformer, Neural ordinary differential equations, Multi-mo-
Keywords dality modeling

Data Availability Private
Global datasets primarily depend on the publicly available ERA5 reanal-
ysis data. Regional data are derived from the publicly accessible data of
Japan's Himawari satellite. However, regional weather observations and
regional reanalysis data are proprietary and cannot be publicly disclosed.

Metadata (Type of Numerical tensors, images
Data)

Model Training and Utilized a Siamese MAE masked autoencoder based pretraining strategy
Fine-Tuning to provide better initialization and learn robust feature representations
hidden within highly volatile weather data. Based on the self-developed
global weather model, multi-modal weather data were cooperated in the
regional weather model to enable rapid refresh accurate weather forecast.

156