How to get another days .nc file from EC?

[Hank0626]

I would like to know how to preprocess the data from zarr file downloaded from WenterBench2.

[andrewlkd]

Indeed you may wish to use relevant datasets listed listed at https://weatherbench2.readthedocs.io/en/latest/data-guide.html

As mentioned in https://github.com/google-deepmind/graphcast/issues/112#issuecomment-2558115672, 1 degree datasets are just the 0.25 degree data subsampled to 1 every 4 points along each of the spatial axes.

As specified in relevant demo notebooks, some transformations were done from the base datasets:

GenCast:

• We accumulated precipitation over 12 hours instead of the default 1 hour.
• For HRES-fc0 sea surface temperature, we assigned NaNs to grid cells in which sea surface temperature was NaN in the ERA5 dataset (this remains fixed at all times).

GraphCast:

• We accumulated precipitation over 6 hours instead of the default 1 hour.
• For HRES data, each time step corresponds to the HRES forecast at leadtime 0, essentially providing an "initialisation" from HRES. See HRES-fc0 in the GraphCast paper for further description. Note that a 6h accumulation of precipitation is not available from HRES, so our model taking HRES inputs does not depend on precipitation. However, because our models predict precipitation, we include the ERA5 precipitation in the example data so it can serve as an illustrative example of ground truth.
• We include ERA5 toa_incident_solar_radiation in the data. Our model uses the radiation at -6h, 0h and +6h as a forcing term for each 1-step prediction. If the radiation is missing from the data (e.g. in an operational setting), it will be computed using a custom implementation that produces values similar to those in ERA5.

[v-weh]

@andrewlkd Would like to ask if precipitation inputs for Gencast will be set to NaN as in the case for Graphcast? or do both models actually take in different inputs?

[andrewlkd]

GenCast doesn't take precipitation as input. It is only a target variable.

[kezhoulumelody]

Hi, are you using 6 hourly-mean data (0:00:00-06:00:00, 06:00:00-12:00:00, ...) to do train and evaluate models? Or are we supposed to just use hourly data sampled from 24 hours (i.e., 0:00:00., 06:00:00, 12:00:00, ...)?

I understand that precipitation is 6-hrly accumulation. Thanks!

[andrewlkd]

We subsample 1 in every 6 frames when using 6-hourly data, so indeed "0:00:00., 06:00:00, 12:00:00, ...".