No experience, so I'm just spitballing here. If the temperatures differ from tank to tank but are a known continuous value, could you just include temperature as a covariate in the model?
Maybe something like:
proc mixed;
class B;
model response=B temperature B*temperature;
run;
Comparisons of means for the levels of B can be controlled by using the AT= option in an LSMEANS statement. If you suspect heterogeneity of response due to temperature, then this may get more complicated, but I would think this would be appropriate as a start.
Steve Denham