As Steve suggested, since you are not using long range electrostatic interactions, the energy does not get conserved. I was having the same problem some time ago when I was doing a test run with only cut off electrostatic potential. I was using nve (with no temperature control). In my case, the energy was sort of constant but the temperature was increasing continously.
One of the things that you can do is to keep track of how much energy is being used in your temp/rescale command as a function of time. If that energy (cummulative) keeps on increasing, it could suggest that system needs to be equilibrated before doing anything else (using other temperature controlled fixes).
One other thing you could do is to use langevin thermostat. It worked out for me, however it would modify your forces (due to random force term).
On the side note, why aren’t you using long range interactions?