search

google-research / weatherbench2

A benchmark for the next generation of data-driven global weather models.
https://weatherbench2.readthedocs.io
Apache License 2.0
400 stars 40 forks source link

Feature request: compare with IFS at higher spatial resolution of 0.1deg #96

Open The-Fonz opened 9 months ago

The-Fonz commented 9 months ago

In renewables forecasting, we care about wind speed, solar irradiation at the location of single wind/solar parks. ECMWF IFS with its 0.1deg resolution gives us good forecast performance, but anything at 0.25deg or less like GFS is far behind.

It would help us a lot if a comparison to ECMWF IFS can be done at the IFS' "native"* resolution of 0.1 lon/lat degrees. The other models can be bilinearly interpolated to match this grid. This would allow us to keep track of these AI models to see "are they useful for renewables forecasting yet".

*The IFS of course has a reduced Gaussian grid so technically the resolution is not really a regular grid of 0.1deg, but practically that's the highest level of detail you get.

raspstephan commented 9 months ago

Hey, that is a great point. We do have the raw high-res IFS HRES data but haven't converted it to Zarr yet. This has already been on my to do list for a while. I can't promise though when I will get to it, so stay tuned for now

jacobbieker commented 6 months ago

Just chiming in that we'd be quite interested in this too! Both for the renewable forecasting case, and since ML forecasting models are now coming out forecasting at the higher resolution, primarily FengWu GHR which seems to forecast on the higher resolution grid.

raspstephan commented 6 months ago

We are currently discussing the best strategy to regrid the data :) But could be a few months before we share it publicly.