Page 193 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Category Climate Change/Natural Disaster
Testbeds or Pilot Electric Transmission System Operators and Distribution System Oper- Change 4.2-Climate
Deployments ators in China
Weather forecasting system for State Grid Shandong Electric Power
Co.LTD.

2 Use Case Description

2�1 Description

Meteorological forecasting plays a crucial role in addressing climate change by delivering
precise weather data that align with industrial needs, particularly in the burgeoning field of
renewable energy. However, existing weather forecasts often fall short of meeting the specific
requirements of different industries. Traditionally, weather forecasting relies on numerical
weather prediction methods, utilizing mathematical and physical equations to simulate
atmospheric and oceanic movements. This approach establishes an initial weather state and
employs computers to solve the equations for future weather conditions numerically. While
effective, this method is computationally intensive, demanding significant processing power.
Moreover, it has limits in regional forecasting, including infrequent updates and diminishing
accuracy over extended forecast periods.

To address these challenges, this approach harnesses advanced algorithms to transform
traditional weather prediction methods. By leveraging AI technology alongside a self-developed
global meteorological model, a high-precision regional weather forecasting model has been
established. This model integrates diverse local data sources, including regional reanalysis
data, real-time meteorological information, radar images, satellite imagery, etc, to enhance the
precision and details of forecasts. The result is a system capable of delivering weather forecast
with a 1-kilometer grid resolution on an hourly basis, thereby providing specialized predictive
support for various industries. [1]

The regional weather forecast model demonstrates significantly superior performance
compared to ECMWF-IFS, achieving an average improvement of RMSE (Root Mean Squared
Error) of 57.8% on key variables in given region. In terms of the global model alone, it
consistently outperforms IFS, achieving an average improvement of 20.5%. In particular, for
T2M (air temperature measured at a height of 2 meters above the ground), our model Baguan
results in an average enhancement of 26.3%. As the lead time increases, the magnitude of our
improvement becomes even more evident. [2]

In order to produce operational weather forecasts, we schedule regular data acquisition
services for publicly available data and get proprietary data through ftp services. We store
data in the Aliyun OSS system and set up the data-transmission pipeline to the product server
for regularly model inference.

In this case, we utilize a Siamese MAE masked autoencoder based pretraining strategy to
provide better initialization and learn robust feature representations hidden within highly
volatile weather data, enabling precise weather insights. We have also independently

157