Question About Input Architecture for GSP Data

[kwon-encored]

Hi @Sukh-P,
I hope you're doing well!

I finally got the model running, and it's now achieving an NMAE10 close to 7! Still working to bring it down further, hopefully closer to your impressive ~3.

For my region of interest, I've divided it into 59 GSPs and have one full year of data (from Dec 1, 2023, to Dec 31, 2024).

I was curious about how your study structured the input data. In my case, I’ve considered two scenarios for training/testing splits:

1. Using GSPs 1 to 57 for training and validation, while reserving GSPs 58 and 59 for testing.
   

2. Using data from all GSPs but splitting temporally—training and validation with data from January to September, and using October to December for testing.
   

Which of these approaches is more aligned with your study, or do you recommend a different strategy?

[Sukh-P]

Hi @kwon-encored, that's great to hear! So I would recommend the training/validation split be done by time for all GSPs, option 2 you present, since there's a chance otherwise there is some data leakage from other GSPs being trained on a period you are testing in option 1.

One thing to note is that for our models we generally have used ~3 years of data for training and around 1 year for validation, since this way the model gets to see at least a few years of each season in training and is likely to generalise better.

All the best with the ongoing work!