As much as I want to, I can't think of a purpose for this. Would be fun as a Kubernetes scheduler! :D
I could actually imagine situations where you have loads that also follow a schedule -- for example, if you have a ton of data and customers can schedule summary generation/analyses, a lot of them will tend to schedule them on the first of the month (that is, after the last month has ended).
In a case like that with a predictable load, you could boot up a bunch of instances a bit before the end of the last day of a month, then turn them off (to go back to normal loads) a bit into the second day of the new month.
Machine learning could help you pick out other peaks, like say, every Monday or Sunday.
If it's predictable, no need for machine learning at all.
> "Machine learning could help you pick out other peaks, like say, every Monday or Sunday."
Or human can script around it, like you just did.