I have not done this before, but several people have.  As I understand it, you use a large supercell and put the impurity atom into one of the lattice sites.  You use the large supercell to avoid have periodic impurities.  You can then check the local density of states with and without the impurity to determine the energy of the impurity electron.

Hello, a few things:

a) check that the impurity is 'far' from its periodic copies in all 3 dimensions. it could be that the impurity is decoupled from images in the plane orthogonal to the GB, but not 'vertically' along the GB itself. you may give more details about the simulation cell.

b) Accurate relaxation is of course essential, and you should always perform it. There is a twist in this context, namely that for isolated point-like impurities you would compute the formation energy at fixed volume, but one could reasonably ask whether this applies to your GB if your impurity disrupts it a lot, which in effect is the same as saying the defect has such a large effective volume as not to be decoupled from its images. 

c) if your system is insulating, the defect may be charged. that can affect the formation energy significantly depending on the Fermi level position as determined by background doping etc.

d) the difference of formation energies is large but not absurd, something like 0.25 vs 1 eV. the effects I mention can explain it to a large extent.