Things to do

[peterdudfield]

Short term

• decide on the name
• save model to github
• publish psp to pypi, so we can install it here
• check nwp variables are in the right order #22
• Define Test set #16
• Evaluate model on 2022 data #2
• add ci for tests
• publish to pypi #21
• Tidy up readme #4
• Add orientation and tilt #6
• Add MIT license
• Adjust for capacity #3

Medium term things

• try ICON data #5
• Benchmark #27
• Tilt + Orientation Inputs
• Live PV inputs
• Easy scaling for a large number of sites
• Simple Data Vis
• OCF running forecast?
• Train model with more PV dropout
• Use OCF-ML-Metrics repo: https://github.com/openclimatefix/ocf-ml-metrics

[zakwatts]

Would be interesting to add a way for the site to capture different inverter characteristics when live pv data is not provided. For example, the case in which you have an undersized inverter would lead to saturation in PV generation. There could be something in PVLib already that could be used to help with this: HERE

[peterdudfield]

@all-contributors please add @peterdudfield for code

[allcontributors]

@peterdudfield

I've put up a pull request to add @peterdudfield! 🎉

[peterdudfield]

@all-contributors please add @zakwatts for ideas

[allcontributors]

@peterdudfield

I've put up a pull request to add @zakwatts! 🎉

[Ludecan]

Hey guys, good evening, this Pablo from Tryolabs, I'm a software engineer with an ML background and I also happened to work at Uruguay's National Weather Service for a few years, so some of these things ring a bell (but no formal training on the physics).

I've been looking for opportunities for collaboration with your project and came up with a few questions and a few ideas for your consideration:

Questions:

• What do you feel is the biggest opportunity where we could land impact?
• What’s the main use of the tool? Is it for determining suitable sites for solar installations? What’s an acceptable level of accuracy?
• How does solar radiation translate to installation capacity? Are there any basics (ideally for beginners 🙂) we can read from the physics?
• Would it be interesting to create a map/sequence of maps of forecasted power generations?

Possible suggestions (to be prioritized):

Features

• Use delta/lagged temporal features
• Use spatial deltas/lags (for incorporating cloud trajectory data)
• Include coordinates in forecast
• Incorporate extraterrestrial solar radiation for the location and point in time? (if I remember this right, we could use this as an upper bound on the solar radiation at the site, ie before entering the atmosphere, it can be approximated deterministically and could help understand yearly/location cycles)
• Consider atmospheric pressure (spatio delta), dew point (for frosting effect) and humidity (to detect fog) for the forecast? The latter two are unlikely to happen on solar panel sites but it shouldn’t be difficult to consider the variables and take them from GFS

Modelling

• AutoGluonTS. Ensemble models for time series forecasting. Could be a good challenger model, computationally intensive but can be tuned for accuraccy/performance tradeoff.
• TSPP. Alternative platform for time series forecasting and validation.
• Error is U-shaped wrt lead forecast time, that's a bit weird. Could be worth investigating

Visualization

• We should be able to create forecast maps or sequences of maps (animations) for a region of interest. Would this be useful?
• Complimentarily we could have maps of confidence intervals for forecast errors

Code

• Main repo uses pip while training repo uses poetry. Would it be useful to switch both to poetry?
• Would it help to configure code styling tools? We use black formatter, isort, and flake8 for linting and have some template configs and Github Actions that could be used for unifying (and automating) code styling.

Looking at the list of issues in your repo, I think an AutoGluonTS Challenger model could be a quick win if we get to improve on the current model.
Do you think it would be worth it for me to implement it into the pipeline?

[peterdudfield]

Hi @Ludecan

Thanks so much for getting in contact, and for lots of your ideas.

Questions (hopefully the answers)

1. I think think the biggest improvements that can be made is to use as many NWPs as possible. For example train on NWP from the UK met office. But we could include ICON, GFS and lots of others.

2. Becasue this produces a live forecast, we see it as being used for live applications, like should I charge my EV today or tomorrow? for example

3. Check out pvlib, this is a bit more physics based.

4. a map of a sequests would be great!

Features

Use delta/lagged temporal features

• This is always useful. Not that the NWPs are for the forecasted time, but still delta features are still useful.

Use spatial deltas/lags (for incorporating cloud trajectory data)

• similar to above

Include coordinates in forecast

• yea that would be handy, currently we put the nwp or sun features from that location, but we could also add the location in too. See here for more features we do use - Update readme.md with more features we use #34

Incorporate extraterrestrial solar radiation for the location and point in time? (if I remember this right, we could use this as an upper bound on the solar radiation at the site, ie before entering the atmosphere, it can be approximated deterministically and could help understand yearly/location cycles)

• yea, we do this already, but its definitely useful

Consider atmospheric pressure (spatio delta), dew point (for frosting effect) and humidity (to detect fog) for the forecast? The latter two are unlikely to happen on solar panel sites but it shouldn’t be difficult to consider the variables and take them from GFS

• pressure is something we should definitely add.

Code

Yea, im in two minds, we are trying to keep this as simple as possible, and I thought requirements would be enough. Might be worth pinning the versions though.

yea we should use those, there are some similar ones in pv-site-predictions which might be worth adding in

Training

• That would be great to use AutoGluonTS. There's probabaly some different options here. You can either add it to pv-site-predictions. You might need to use upgrade it to train on ICON NWP data as this is public. The model we have trained on is using Metoffice UKV data. Another option is to build a seperate model training code. This could be in this repo or somewhere else. Then we can compare the models e.t.c. What do you think?

[Ludecan]

Hey @peterdudfield, thank you very much for your answers. Very helpful to start gaining context.

I was thinking on adding AutoGluonTS as a configuration option for the regressor of the existing pipeline. That way we can reuse all the work you guys have done for other components.
I think for that I would need to implement a new regressor class in pv-site-predictions/psp/models/regressorsright?
And then create a config for the training script to use the new regressor

And regarding the data, I followed the links to this dataset. Is this the UKV data? If so maybe I can start working with it and then in a separate step add in the ICON NWP data. What do you think?

[peterdudfield]

Yea that is about right.

That dataset is the PV data, so not the NWP. Unfortunately the UKV NWP data is not public. But you do need the that PV data.

Yea please have a go at adding that regressor, and myself and/or @zakwatts could help you out

[peterdudfield]

Hey @peterdudfield, thank you very much for your answers. Very helpful to start gaining context.

Yes, I was thinking on adding AutoGluonTS as a configuration option for the regressor of the existing pipeline. That way we can reuse all the work you guys have done for other components. I think for that I would need to implement a new regressor class in pv-site-predictions/psp/models/regressorsright? And then create a config for the training script to use the new regressor

And regarding the data, I followed the links to this dataset. Is this the UKV data? If so maybe I can start working with it and then in a separate step add in the ICON NWP data. What do you think?

Hi @Ludecan, see my answer above. This message from you seemed to come twice so I'm a little b confused

[Ludecan]

Got it. Will start working on that and keep you posted. Sorry for the double post, just deleted the duplicate.

[peterdudfield]

No worries, thanks in advance @Ludecan