This would entail a bigger change to PVNet as instead of a single forward pass creating a full forecast, this would entail a different forward pass per forecast step (or chunk of forecast steps), and is somewhat similar to the time step trick used for MetNet, but could help with the forecasts. Also not sure how we'll or if LoRA works for non transformers.
In PanguWu GHR they use LoRA to correct per forecast step biases, to pretty great effect. https://github.com/openclimatefix/graph_weather/issues/84
Detailed Description
This would entail a bigger change to PVNet as instead of a single forward pass creating a full forecast, this would entail a different forward pass per forecast step (or chunk of forecast steps), and is somewhat similar to the time step trick used for MetNet, but could help with the forecasts. Also not sure how we'll or if LoRA works for non transformers.
Context
Better forecasts are good
Possible Implementation