Use minimal data types for PV

[JackKelly]

This is what a PV batch looks like in v15. All the 64-bit data types could be 32-bit:

[peterdudfield]

Ill have a look at v16, just incase its been done already

[peterdudfield]

Looking in /mnt/storage_ssd_4tb/data/ocf/solar_pv_nowcasting/nowcasting_dataset_pipeline/prepared_ML_training_data/v16/test/pv

import xarray as xr
pv = xr.load_dataset('000000.nc',engine='h5netcdf')

so the varibales that need upgrading are

• x_coords -> float32
• y_coords -> float32
• id -> int32
• example_index
• id_index
• time_index

[JackKelly]

sounds good!

I think we need to use float32 for id (so we can use an ID value of NaN to represent "missing" PV systems)

Also, I think we can completely throw away example_index, id_index, and time_index because they're just repeats of the example, id, and time dimensions 🙂 Please see issues #629 and #630 for more context on throwing away example_index, id_index, and time_index.