Inputs: recent history of PV data and NWP, PV angle & azimuth, Sun angle and azimuth (or just panel effective irradiance), and maybe local PV systems (although that could be a separate model. Some people won't have access to neighbouring PV data)
Can be used as is. And the model will try to infer things like soiling from recent PV history. But much better to fine tune on individual PV system.
Train on real PV data and pvlib simulation
Maybe make it "temporal resolution independent" (i.e. don't use an RNN. Just use self attention?)
No satellite imagery. Keep it simple.
Inputs: recent history of PV data and NWP, PV angle & azimuth, Sun angle and azimuth (or just panel effective irradiance), and maybe local PV systems (although that could be a separate model. Some people won't have access to neighbouring PV data)
Can be used as is. And the model will try to infer things like soiling from recent PV history. But much better to fine tune on individual PV system.
Train on real PV data and pvlib simulation
Maybe make it "temporal resolution independent" (i.e. don't use an RNN. Just use self attention?)