[META] forecast 8 hrs into the future

[JackKelly]

• Try predicting 8 hours of GSP. Only change GSP and NWP forecast horizons, nothing else. That should "magically work" because it's only the satellite data source that cares about the Sun's angle, and so we should still get the same t0s, if the satellite data source's history duration and forecast duration are the same.
• Fix bug when trying to predict 8 hrs into the future. #138
• Fix plotting
• Include Sun's angle & azimuth in GSP predictions
• Try removing the RNN? #135
• always downsample PV to half hourly to keep things simple for now? #140
• The t0_idx should be specified by the DataLoader, and modified by anything that resamples, and passed into the model. #141
• Make satellite optional
• Resample NWP to half hourly #143
• Modify the NWP data loader so it uses most recent init time for history. #142
• Implement tweaked model which predicts single timeseries at once #139
• Give the model more GSP history and NWP history. (And change T0_IDX_30_MIN)
• Solar azimuth and elevation should be computed more like how we compute time fourier features: e.g. we give it a list of BatchKeys to create solar azimuth and elevation for.
• Change the evaluation metrics to only compute during daylight hours
• Give the model all of yesterday's PV & GSP & NWP history #134

[JackKelly]

Maybe the simplest thing to do is to not filter by the Sun's angle at all, and maybe extend the forecast horizon to, like, 24 hours so every example is guaranteed to have some daylight and nighttime. Although, of course, the HRV imagery won't show anything useful at night!

So each example could start at literally any time of day (00:00, 06:55, 12:00, etc.)

[JackKelly]

Any thoughts about the idea in the comment immediately above this comment, @peterdudfield and @jacobbieker?

[peterdudfield]

I think its not crazy, to just do full 24 hours start time. Could always look into it if it really starts effecting performance?
But as a good default, probably worth trying

[peterdudfield]

The your MAE will probably go down? easier to predict at night. Have to think how we compare that properly

[peterdudfield]

I would stick to a 8 hour horizon at the moment though, just 24 hours start time.

let's get that working first, and then we can think about extending

[JackKelly]

All sounds good!

And, yeah, I'd need to make sure the validation metrics are computed in the same way (using only daylight hours).

[jacobbieker]

Sorry for being a bit delayed with this, but yeah, agree with peter, just have it start at any time and the 8 hour horizon,