I also have a sneaky suspicion that there is an auxillary genset hidden in an innocuous white cabin 10 metres away from the SpC compound at the Badgers Farm (UK) site. (Can't see what else would require massive wires running from this cabin over a wall out of sight, then back into the SpC cabinets).
I know nothing about the Badgers Farm site. However, there is one other obvious candidate - a 400 kWh battery pallet. Tesla has been slowly rolling them out to Supercharger locations, starting with areas that have large demand charges or large time of use swings - the battery pack lets them level the very peaky load.
Senior Tesla folks like Straubel have made presentations about the packs and there usage a few times in the last few years - it's apparently pretty much the guts of five 85 kWh battery packs arranged a little differently and places on a 4' x 6' pallet.
Walter
- - - Updated - - -
Good point, hadn't considered that. The best choice is probably just to build a model with a feature set that includes things like SOC, occupancy, power output, average arrivals for the period, etc. It wouldn't be exactly right, but remember that all models are wrong but some models are useful. I'd be happy with something more useful than any of those items by themselves.
That sounds like a very complicated model to try to guess all of those factors, and very error prone. I think the only way to really address this is from the cars up, instead - continue improving on the energy usage modeling and prediction, fix the navigation routing system.
The car knows from the energy usage on the previous leg how far off of norms it is (from extra people & luggage, or ski racks, or weather effects - though the weather might be part of the usage model itself.) The car also knows what sort of charge taper it will have in the session, based on battery temperatures, etc.
That means each car knows how long it will need at a unpaired SpC station at each stop (and can take a reasonable guess at how long it'll need at a paired station.)
In my opinion, the only way to make a really practical usage model and predictor is to combine these data sets - have each car pass a predicted time of arrival and time to charge at each charger back to a server at HQ. This server will then know hours ahead about potential overloading (and I've suggested before, if there's enough overlap in the system, it could work cooperatively with later cars to shift them on to less loaded nodes.)
Obviously, there's still some room for error here - people don't necessarily leave the moment the car is ready, and drivers may not hit their ETA (though that's really just another opportunity for deep learning in the system about the real road net and driving patterns.) This approach will also only work for folks using Navigation or driving the same trip the same way repeatedly (Tesla already tries to learn your commute, right?) On the whole, though, I think it's the only way to get a reliable model.
Walter
