Train GSP PVNet 36 hours

[peterdudfield]

Train PVnet out to 36 hours.

Would be interesting to see the results compared to just training a model out to 8 hours.

We might have to just trained to 33 hours, due to some data delays in live

[Sukh-P]

Started on this now, was slightly delayed as we were waiting for more NWP data with longer forecast horizons to be downloaded. For the first pass will be using UKV NWP data which goes out to 36 hours and ECMWF to 84 hours as well as satellite data to produce 36 hour PV forecasts for GSPs, will be good to compare to previous experiments that have been ran on extending the forecast horizon of PVNet to 36 hours

[Sukh-P]

Some initial results from training PVNet can be found here and will also be added as a markdown to the PVNet repo

[dfulu]

Some initial results from training PVNet can be found here and will also be added as a markdown to the PVNet repo

Personally I think having markdowns inside the repo is untidy and maintaining google docs notes on the experiments which are already linked to from inside the repo is a cleaner way of tracking experiments.

Keeping it in google docs means:

• We don't have to go through testing and merge requests just for making notes or changes to notes on an experiment, so are more maintainable
• The experimental notes are in a richer format and can be discussed via comments
• The PVNet version number is linked to changes in the code rather than also being linked to changes in experimental notes

But I know that @peterdudfield has a different opinion and neither of us can be convinced by the other 😆

[Sukh-P]

FYI I have added the day ahead model here https://huggingface.co/openclimatefix/pvnet_uk_region_day_ahead/tree/main, will close this for now as we will be working on deploying it