In my case I've just been looking at an atmospheric boundary layer through an empty domain, so the mean is pretty well defined. I just start the averaging at a certain point in time after the simulation has had time to start up, and let it average over the rest of the run time