[ML Idea] Simple PV power predictions using single timestep of satellite data

[JackKelly]

[JackKelly]

Made a start on exploring how to load individual PV systems here: https://github.com/openclimatefix/power_perceiver/blob/main/notebooks/2022-03-09_explore_loading_single_pv_system/explore_loading_single_pv_system.ipynb

Update: Some more thoughts on experiments to run:

• What happens if we give the model a rectangular crop of the satellite imagery (as per #18)
• Which satellite channels are most informative?
• Explore if performance changes when doing single-pv-system-at-a-time versus a Perceiver (IO?) model which predicts multiple PV systems at once (probably have to use a different spatial encoding). Specifically:
  • Try single PV system using "normal" spatial encoding.
  • Then try multiple PV systems using normal spatial encoding
  • If it doesn't work as well, then try helping the model. e.g.
    • The Sun's rays are parallel. So use a rotated position encoding, whereby the new y points towards the Sun, and the new x is perpendicular to the new y. Then, to find the pixels between a PV system and the Sun, the ML model just needs to find all the pixels with the same x coordinate as the PV system.
    • Older ideas, which are probably worse and might not be worth trying:
      • Each query element is for a single PV system. Include the location of all the pixels in the "line of sight" between that PV system and the Sun.
      • Or: Get the model to explicitly predict the line of pixels from the PV system to the Sun. e.g. it predicts a sparse image of the line between the PV system and Sun.

The longer-term idea is to explore whether we need to predict each PV system individually, or can we predict multiple PV systems in one go (which would be helpful when we're combining multiple modalities.

[JackKelly]

Next step: Build a pytorch dataloader which loads just a single PV system at a time