Add flexibility of Input Data Sources at Inference time

[jacobbieker]

The Clay EO Foundation Model does some interesting tricks to handle various inputs and missing/new inputs at run time compared to PVNet's approach of being somewhat flexible, but needing to be retrained when new sources are added. Primarily, for image inputs, it uses https://github.com/zhu-xlab/DOFA to be able to take in arbitrary image bands, both in number and wavelength, and process them well. This could be very useful for transfer learning for PVNet from Europe to India, or anywhere else, where other satellite imagery might be available (i.e. India's geostationary satellite, or JAXA's) compared to EUMETSAT, or to more flexibly include HRV, or other, more transient imagery from non-geostationary satellites.

Detailed Description

Take some of the ideas from DOFA and Clay to make PVNet more generalizable and flexible. Ideally, this would mean if data is missing, the model would run more fine without it than it currently does. This might be able to be extended to non-imagery as well, like NWPs, although not sure how well that would work exactly.

Possible Implementation

DOFA and Clay repos both have implementations of the idea