Page 244 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Metadata (Type of 1) Meteorological data including time-series measurements (e.g., wind
Data) speed, solar irradiance, temperature, humidity) and geospatial data
(e.g., longitude-latitude coordinates).
2) Data formats include NetCDF for large-scale meteorological datasets
and CSV for site-specific records.

Metadata includes attributes such as timestamp, geographic coordi-
nates (latitude, longitude), sensor type, measurement unit, and data
collection frequency (e.g., 15-minute intervals).

Model Training and The AI Large Model, based on Huawei’s Pangu model [1], is pre-trained
Fine-Tuning on global meteorological datasets and fine-tuned using transfer learn-
ing on proprietary datasets from China Huadian and Beijing Jiutian.
Fine-tuning involves supervised learning with historical meteorological
and power generation data to optimize predictions for wind speed,
solar irradiance, and power output. No Retrieval-Augmented Gener-
ation (RAG) or external knowledge base is used; the model relies on
end-to-end training with domain-specific data.

Testbeds or Pilot Pilot verification at Huadian Jiangsu pilot station, reducing assessment
Deployments fines in the first half of 2024. Based on Huawei high-performance AI
computing ability, including training and inference ability, Our solu-
tion provides different deployment policies in different scenarios. At
Huadian, the AI model runs on a dedicated computing cards and does
not require supercomputer clusters. The entire deployment is efficient
and highly scalable.

2 Use Case Description

2�1 Description

Introduction: Amid the global push for sustainable development, the installed capacity
of renewable energy continues to grow rapidly. However, the inherent unpredictability,
variability, and intermittency of wind and solar power pose major challenges to power grid
dispatching. Meanwhile, both regulatory assessments and electricity market trading demand
higher forecasting precision and efficiency from renewable energy operators—requirements
that traditional numerical weather prediction models struggle to meet. This use case aims to
enhance the accuracy and efficiency of new energy meteorological power forecasts to support
grid stability, optimize electricity trading, and reduce assessment penalties.

Solution Overview: China Huadian, in collaboration with Huawei, developed a new energy
meteorological power prediction solution using a cloud-edge collaboration three-layer
architecture:

• Group meteorological model: Processes large-scale historical and real-time
meteorological data.
• Regional micro meteorological engine: Refines forecasts for specific regions.
• Site edge power prediction service system: Delivers localized, real-time predictions.
• The solution utilizes AI Large Model technology (e.g., Huawei’s Pangu model), which
outperforms traditional numerical models by completing global high-resolution forecasts
in seconds, nearly 10,000 times faster. This approach achieves forecast accuracies of

208