Download as odt, pdf, or txt
Download as odt, pdf, or txt
You are on page 1of 1

Graphcast for weather forecast

1. Advantage compare with traditional method


• Traditional method use Numerical Weather Prediction (NWP) which begins with
carefully define physics equation → translate into algorithm and run and supercomputer.
• Deep learning use data instead of physical equations → Can recognize some pattern the
physical equation cannot capture.
• Fast predict, output the result in < 1 minute (Google TPU v4) compare with >= 10-day
forecast using HRES.
• More accurate than HRES method accord Science paper.

2. Model
2.1 Input – output
• Gridded datatype – store with GRIB file format: The 2D datatype base on 5 unique
attributes
◦ Valid time
◦ Level
◦ Parameter
◦ Origin
◦ Navigation
• Need only data of 6 hours ago and current state of weather → Output the weather of 6
hours after. Can apply regression to forecast up to 10 day.

You might also like