New features

[peterdudfield]

• [x] Investigate results form live
• [x] resample every forecast run to 30 mins intervals. Less likely to have spiky results - https://github.com/openclimatefix/gradboost_pv/pull/46
• [x] remove sde from the training - https://github.com/openclimatefix/gradboost_pv/pull/48
• [x] add more GSP LAG features - https://github.com/openclimatefix/gradboost_pv/pull/49
• [x] remove hcct from the training - https://github.com/openclimatefix/gradboost_pv/pull/51
• [ ] add mcc and hcc to nwp variables - https://github.com/openclimatefix/gradboost_pv/pull/52
• [x] add more n_estimators - https://github.com/openclimatefix/gradboost_pv/pull/53
• [ ] add smoother to results
• [ ] extended to 40 hours. This would use the most recent NWP data
• [ ] Look into using deeper regression trees
• [ ] Use historic NWP data, not just forecasts
• [x] Normalize features before training (dont do, see comment below)

Would be interested to think what people think I should do first? @JackKelly @jacobbieker @dantravers

[jacobbieker]

I would probably start with removing the sde from training, and then probably more lag features? I think XGBoost models don't need the data to be normalized, so not sure that's necessary, although I guess if the units are different between CEDA and live MetOffice it probably makes sense to do that first.

[peterdudfield]

Bonus one is to add mcc and hcc to nwp variables

[JackKelly]

if the units are different between CEDA and live MetOffice it probably makes sense to do that first

Yeah, it's pretty essential that the data the model sees at inference time is exactly the same as the data it sees at training time :slightly_smiling_face: so I agree that sounds like the priority!

And I agree with @jacobbieker that I don't think XGBoost models require the data to be normalised (because it chops real-valued inputs up into bins).

Does the model also get historical NWP data? If not, I think that might help a bit: i.e. if the model gets lagged GSP data for n timesteps in the past, then it might be useful to give the model NWP data for those same timesteps so the model can see the difference between the expected forecast (given the NWP) and what actually happened in the recent past. But maybe the model is already doing that?

[peterdudfield]

Thanks @JackKelly and @jacobbieker , i re-ordered above, do you that order is about right?

[JackKelly]

Lgtm!

[jacobbieker]

Looks great!

[peterdudfield]

Thanks, @dantravers you happy with this?

[dantravers]

Looks reasonable to me! I'd be curious to see if this does well, so could be higher? Use historic NWP data, not just forecasts But seems sensible. Thanks for asking the open question!