Closed jacobbieker closed 1 year ago
Also, so the model can learn how to deal with missing satellite data, have every 5 or 10% of training examples be with zeroe'd out satellite channels
Could downsample HRV to the non-HRV channel sizes before putting into the model? Or the opposite, and upsize the non-HRV to the HRV before the whole downsample step in MetNet
In tech meeting, having this forecast the next 24 hours seems like the way to go, in which case I think MetNet-2 would work better as its supposed to be a smaller model and worked better for 12 hour forecasts
Detailed Description
As discussed on tech planning meetings, having a national forecast model might be helpful, and MetNet fits the bill with the ability to deal with very large context images.
Context
Improving the national forecast is good
Possible Implementation
The code for both MetNet-2 and MetNet is arleady setup for NWP and Satellite, and topographic inputs, as long as they are the same resolution. The trickier bit is including PV and GSP data. One way would be to just put them at the pixels they contain in the overall image as more channels? Would probably be the simplest, GSP level would be one channel with the values being the PV generation for that timestep for each pixel containing the GSP. PV would be similar, although for piuxels with multipl PV systems, could take the average of them? Or have multiple channels for PV so every PV system is included.
Other part needed for this is to implement the new data loader with data pipes, probably.