GRIB files output switched from Kelvin to Celsius starting on run 2024-04-17 0z

[patrick-zippenfenig]

GRIB files on AWS switched from Kelvin to Celsius on run 2024-04-17 0z and current runs after 2024-04-18 12z. Record parameterUnits shows Kelvin, but Celsius is stored. Effects 2m and pressure level temperature.

Originally reported by @bzigterman in https://github.com/open-meteo/open-meteo/issues/748

[patrick-zippenfenig]

I applied a simple correction by -273.15. For most locations this corrects the issue. However, some locations show unreasonable differences of 15K compared to GFS. I suspect that issue could already be present in the initialisation data for GFS GraphCast.

You can try it here: https://open-meteo.com/en/docs#latitude=40.1106&longitude=-88.2073&current=&hourly=temperature_2m,temperature_500hPa&past_days=5&models=gfs_global,gfs_graphcast025

Edit: The 15K issue only effected 2 runs. Others are fine. This is just an observation

[SadeghTabas-NOAA]

@patrick-zippenfenig Thanks for notifying us about this issue, we will fix the issue and upload the forecasts again ASAP.

[SadeghTabas-NOAA]

@patrick-zippenfenig I was wondering if the issue is with either 13PL or 37PL forecasts? Also, I wanted to let you know that we will soon shift to GraphCastGFS V2 which will have a higher accuracy for 13PLs. Thanks

[patrick-zippenfenig]

I am using 37PL. Will you only use the new weights for 13PL?

[SadeghTabas-NOAA]

@patrick-zippenfenig Yes, at this time we are only focused on 13PL, we might stop providing forecasts for 37PL in a couple of days. Please let me know if you can see the same issue with 13PL as well, so I can work on it. Also, I wanted to note that the 13PL forecasts are more accurate compared to 37PL in the current version.

[patrick-zippenfenig]

Oh, that's good to know. I will switch to 13PL and disable the automatic kelvin/celsius detection to see if there are any issues

[SadeghTabas-NOAA]

Thank you so much

[SadeghTabas-NOAA]

@patrick-zippenfenig I'm checking the 37PL forecasts for 2024-04-17 0z and I dont see the issue for 2mt variable, I checked .006 and .036 files. Could you please let me know which file has the issue? Thanks

[patrick-zippenfenig]

Oh sorry, I believe it is the 37PL 2024-04-17 12z run. I am using the ECMWF eccodes library to read GRIB2 files. I hope it is easily reproducible and not dependant on some library/environment combination

[SadeghTabas-NOAA]

@patrick-zippenfenig Thanks for providing this. The corrected forecasts would be available in 2 or 3 hours. The model ICs were corrupted for those dates which resulted issues in all variables including TMP. Thanks

[SadeghTabas-NOAA]

As I fixed the forecasts, I close this issue.

[patrick-zippenfenig]

Thank you!