Embedding of GSP ID and PV System ID

openclimatefix / nowcasting_dataloader

PyTorch Dataloader for working with multi-modal data for nowcasting applications

https://nowcasting-dataloader.readthedocs.io/

MIT License

7 stars 0 forks source link

Embedding of GSP ID and PV System ID #21

Open JackKelly opened 2 years ago

JackKelly commented 2 years ago

Detailed Description

Let's experiment with giving our models:

An embedding of the GSP ID (used in the query? And for each row of GSP-level PV data if we use historical GSP-level PV?)
An embedding of the PV System ID (could be used as pat of the position encoding for each row of PV data)

JackKelly commented 2 years ago

This implemented now, I think. Is that right, @jacobbieker?

jacobbieker commented 2 years ago

Yeah, we give it the ID, but we don't actually do anything with the ID. AS in, if a systems ID is 576, then the model gets 576 as the input, if that works?

JackKelly commented 2 years ago

Yeah... although it might be slightly better to first wrap the ID in an embedding (which should "magically" cluster similar GSPs in vector space... and should help the model understand that GSP ID 100 isn't necessarily "similar" to ID 101, even though they have adjacent numerical IDs)

jacobbieker commented 2 years ago

Yeah, that sounds good!

JackKelly commented 2 years ago

yeah, it might help a bit... not sure though!