[x] Get old model working again, which infers PV power for 8 PV systems from a single timestep of satellite imagery
[x] Add in time_utc_fourier encoding for imagery and queries
[x] Predict GSP PV power for the following GSP timestep (e.g. if the imagery is at 11:20am then predict GSP power at 11:30am).
[x] Concatenate multiple timesteps of imagery. e.g. imagery at 11:00, 11:15, 11:30, and 11:45am, to predict GSP at 11:30am and predict PV power at the last timestep
[ ] #51
[ ] fix bug: there's a problem when a PV system appears multiple times in an example: we're giving the PV system too much weight. Maybe mask the loss on duplicate PV systems. Or mask the attention?
[ ] #43
[x] Use one (and the same) embedding for PV ID and GSP ID (making sure they don't clash!) Maybe it'll figure out which PV IDs are associated with which GSPs?
[x] Plot satellite images to wandb. Also plot location of GSP and PV systems.
[ ] Mixture density network (at least for GSP PV Power)
[x] turn off watch command for wandb
[ ] Try different number of images
[ ] Try Perceiver IO which sees the images as separate inputs
[x] Send hyperparams to wandb
[x] learning rate scheduler? LR warmup then drop like this and also see Stack Overflow
[ ] Randomly change the timestep of the PV prediction to any timestep covered by the sequence of images.
[x] Absolute spatial location
[x] Try predicting multiple timesteps of GSP or PV at once
time_utc_fourier
encoding for imagery and queriesnan_to_num
for the topographical data