Experiment with dynamic fusion architecture

[felix-e-h-p]

Pick up relating to PVNet updates

[Sharkyii]

Hi, I’m Sneh an enthusiastic ML engineer, and I’d love to start contributing. I’ve gone through the repo, and this issue seems like a great entry point, but I wanted to confirm that it’s appropriate to begin here.

My plan is:

• Implement a dynamic fusion layer that learns per-modality weights (NWP, satellite, PV history, sun, location).
• Add a new dynamic_fusion model config and run the same experiments as the current late_fusion setup.
• Compare metrics and include a brief summary of the results in the PR.

Does this direction align with what you had in mind, and do you have a preference between a gating-based or attention-based fusion design?
If possible, could you please assign the issue to me @felix-e-h-p ?

Thanks!

[AUdaltsova]

Hi @Sharkyii, thanks for offering to help! Unfortunately, this is an internal issue and is not open to open source contributors. Hope something else catches your eye!