I tried to fit a physical model (rather than fitting to EPA coefficients).
Here's a fit of my 3 datapoints, plus two datapoints
I made up for 25mph & 35mph (though they seem reasonable based on an aero model, in conjunction with reducing fixed losses between 25mph & 45mph).
Wolfram|Alpha: Making the world’s knowledge computable
This is a formula for Watts, NOT Wh/mi. But the y-axis data points are entered as Wh/mi * velocity. This gives you the actual watts for a given speed. So you can try to fit your own data if you want. Warning: small changes in the data can lead to very large changes in coefficients...
You can divide through by velocity to get the Wh/mi formula.
So it's watts = 0.035*v^3 + 101.5*v + 362 (dropped the minimal v^2 term, which I can't think of a physical reason for - it may well exist though)
Or Wh/mi = 0.035*v^2 + 101.5 + 362/v
Notably, 362 (units of watts) is the fixed loss when stationary. I basically picked this value to fit my three points...so don't read too much into it - though I think it's in the general ballpark.
101.5 (units of Wh/mi) are the fixed losses (rolling resistance, etc.). Again, don't read much into this. This is the one that is going to change quite a bit when you put on different tires.
Anyway, with some very careful data gathering of the 25mph & 35mph plots it might be possible to get a better model. I think the 0.035 coefficient is actually pretty reasonable, might be slightly high. This coefficient would change when you put aeros on. The other coefficients potentially could be quite off (fixed losses could be lower and rolling resistance higher, or vice versa), and are subject to errors in the datapoints used to fit the model.
To validate this model as being "good", I'd need to get 138Wh/mi at 25mph and 155Wh/mi at 35mph (I'm not sure how probable this really is). So please don't take anything here as "accurate." It's just playing around.
This formula also predicts 303Wh/mi at 75mph, which to me seems high. So the squared term coefficent may be SLIGHTLY high.
Also worth noting that the fit of this model is highly dependent on the input data. So if the meter in the car becomes less accurate at lower consumption (some have claimed this) it would make it very difficult to develop an accurate model, unless you know how that error behaves.