I'm working with a cleaned hourly table of electricity generation by source (MW), indexed by UTC timestamp.

I want to predict total thermal generation from the renewable sources. That means:

- the target is the sum of several thermal series, but it should be missing if any one of them is missing (a missing reading isn't zero)

- the inputs are the renewable series plus hour, day of week and month

- only hours where the target and every input are present should be kept

- the last full year should be held out as a test set, and the two years before it used as validation years in expanding-window folds

I can't split randomly, because neighbouring hours are almost identical and that would leak information. How can I set this up in pandas?