Page 28 - SAMENA Trends - July-August 2023

REGIONAL & MEMBERS UPDATES SAMENA TRENDS

HUAWEI CLOUD Researchers Develop AI Weather Forecast System with

10,000x Faster Predictions Compared to The Traditional Model

HUAWEI CLOUD published a breakthrough
paper on the Pangu Weather AI model in one of
the world's top scientific journals, Nature. The
paper describes how to develop a precise and
accurate global AI weather forecast system
based on deep learning using 43 years of data.
Pangu-Weather is the first AI prediction model
to demonstrate higher precision than traditional
numerical weather forecast methods. The model
allows a 10,000x improvement in prediction
speed, reducing global weather prediction time
to just seconds. Pangu-Weather challenges the
previously held assumptions that the accuracy
of AI weather forecast is inferior to traditional
numerical forecasts. The model, developed
by the HUAWEI CLOUD team, is the first AI
prediction model with higher precision than
traditional numerical prediction methods. The
paper, titled "Accurate medium-range global addresses these challenges During AI team chose to focus on weather
weather forecasting with 3D neural networks" scientific trials, Pangu-Weather model predictions, Dr. Tian Qi, Chief Scientist of
provides independent verifications of these has demonstrated its higher precision HUAWEI CLOUD AI Field, an IEEE Fellow,
capabilities. The publication marks the first compared to traditional numerical and Academician of the International
time that employees of a Chinese technology prediction methods for forecasts of 1 Eurasian Academy of Sciences,
company are the sole authors of a Nature hour to 7 days, with a prediction speed explained "Weather forecasting is one
paper, according to Nature Index. With the rapid gain of 10,000 times. The model can of the most important scenarios in the
development of computing power over the past accurately predict in seconds fine-grained field of scientific computing because
30 years, the accuracy of numerical weather meteorological features including meteorological prediction is a very
forecast has improved dramatically, providing humidity, wind speed, temperature, and complex system, yet it is difficult to
extreme disaster warning and climate change sea level pressure. The model uses a cover all aspects of mathematical and
predictions. But the method remains relatively 3D Earth-Specific Transformer (3DEST) physical knowledge. We are therefore
time-consuming. To improve prediction speeds, architecture to process complex non- delighted that our research has been
researchers have been exploring how to using uniform 3D meteorological data. Using recognized by the Nature magazine.
deep learning methods. Still, the precision of a hierarchical, temporal, aggregation AI models can mine statistical laws of
AI-based forecasting for medium and long-term strategy, the model was trained for atmospheric evolution from massive
forecasts has remained inferior to numerical different forecast intervals using 1 hour, data. At present, Pangu-Weather mainly
forecasts. AI has been mostly unable to 3-hour, 6- hour and 24-hour intervals. completes the work of the forecast
predict extreme and unusual weather such This resulted in a minimization of the system, and its main ability is to predict
as typhoons. Every year, there are around 80 quantity of iterations for predicting a the evolution of atmospheric states. Our
typhoons worldwide. In 2022, in China alone, the meteorological condition at a specific ultimate goal is to build next-generation
direct economic loss caused by typhoons was time and a reduction in erroneous weather forecasting framework using AI
5.42 billion yuan, according to the figures from forecasts. To train the model for specific technologies to strengthen the existing
China Ministry of Emergency Management. The time intervals, the researchers trained forecasting systems. " Commenting
earlier that warnings can be sent out, the easier 100 epochs (cycles) using hourly on the significance and quality of the
and better it is to make adequate preparations. samples of weather data from 1979- research by HUAWEI CLOUD, academic
Because of their speed, AI weather forecast 2021. Each of the sub-models that reviewers from Nature explained that
models have been attractive but have lacked resulted required 16 days of training not only is Pangu-Weather very easy to
precision for two reasons. First, the existing on 192 V100 graphics cards. Pangu- download and run, but that it executed
AI meteorological forecast models are based Weather Model can now complete 24- quickly on even a desktop computer. "This
on 2D neural networks, which cannot process hour global weather forecasts in just means that anyone in the meteorological
uneven 3D meteorological data well. Second, 1.4 seconds on a V100 graphics card, community can now run and test these
medium-range weather forecast can suffer from a 10,000-time improvement compared models to their hearts' desire. What a
cumulative forecast errors when the model is with the traditional numerical prediction. great opportunity for the community
called too many times. How Pangu-Weather Explaining why the HUAWEI CLOUD to explore how well the model predicts

28 JULY-AUGUST 2023