[StargazyPi]

Depends; the input data set could be massive, incorporating things like TV listings data (half-time kettle surges are a huge part of our power consumption profile), weather, publicly listed events, perhaps scraped from websites and so quite dirty.

I can see this transitioning quite easily from a 'we can sanitise this data ourselves' job to a 'screw it, let a massive neural net figure this out' one.

[redcalx]

My initial best guess is that the style of analysis you're suggesting (TV guides, etc), doesn't offer sufficient safety margin in the event that a prediction is wrong. If reducing the safety margin was an option then there are probably lots of things that can be done, but if the goal is to maintain a safety margin akin to the current one then my instinct is that the gains to be had are minimal - and that the interesting work is in power storage (battery tech, cryogenic storage, etc.)