search

openclimatefix / power_perceiver

Machine learning experiments using the Perceiver IO model to forecast the electricity system (starting with solar)
MIT License
7 stars 1 forks source link

Give the model all of yesterday's PV & GSP & NWP history #134

Open JackKelly opened 2 years ago

JackKelly commented 2 years ago

Let's say we create a forecast at 4am.

There won't be any PV data for the last hour (3am - 4am). So how will our model calibrate itself at inference time?

One solution could be that we always give the model the entire PV power generation for "yesterday" (from dawn to dusk).

We probably don't need "yesterday" to be at 5-minutely resolution. Hourly might be fine. So we'd give the model yesterday's NWPs and yesterday's PV and GSP power, all at hourly resolution.

But, maybe we only need to worry about yesterday's GSP power for now (because we're focused on GSP predictions) and we can think about how to provide yesterday's PV power at another time.

Could also try removing the RNN (#135) as life might be simpler if we didn't have an RNN!

JackKelly commented 2 years ago

Yeah, to start with, maybe just give the model more GSP history.