Also, you need people to move. Because you only have power data, you only know the aproximate signal strength if the person isn't moving.
I can see this "technique" being used in an episode of NCIS, CSI or similar cop show. <plot> We need to find this guy, but all we have is his cell number.
Police tech #1 "Sure no problem, I will just connect to his phone and measure the battery drain..... Got him, he is Downtown at 5th and Main- lets go arrest him!!!" </plot>
The "general habits" assumption seems implausible, but it isn't. Let's say you want to track all the citizens of the Bay Area. You know that, in general, ppl from the Bay area are in the Bay Area. Therefore, you map the cell reception in Bay area. You also take note of major transportation routes and patterns. Now, if you want to track a particular target, all you'll need is their power usage.
They learned the typical profiles of power use for different journeys and then tried to distinguish between them (at least in one experiment).
By having a sufficiently accurate map, you could determine the likelihood of a route by the changes in signal. Even without pre-mapping there are some estimations that could be made: a signal that follows a clean sinusoidal curve with a certain rate of change could likely be modeled by rural highway driving.
All I can gather is that they would also need to know what tower you were talking to, and then based on the drain they could probably guess where you were based on some heuristic. Meaning, if you are talking to tower x, and the battery drain is high, you could guess that you are either far from the tower or indoors somewhere. It still seems to me that this is dubious at best. I get that technology is always changing, but wouldn't it just be easier to exploit a security hole?