If what you want to measure is them exactly restrained like that, then I guess you should keep the restraints. You shouldn’t need the Z restraints in for the reference group either. The question is how to compare it to anything observable.
With longer equilibration time, I mean that if you change any parameters (including restraints) between the conformation generation and the umbrella simulations you might need to equilibrate the umbrella windows longer. In my opinion you should always use the -b option to discard the beginning of each simulation. I can’t say how much. It’s probably a bit overkill to discard as much as 5 ns out of a 10 ns simulation. The best way to check is usually to find a time block that gives a reasonably reliable estimate and then analyse using a sliding window to find the time when the simulations are reasonably stable, at least without a trend in the PMF values as you start analysing later. E.g.
-b 0 -e 2500
-b 500 -e 3000
-b 1000 -e 3500
-b 1500 -e 4000
-b 2000 -e 4500
-b 2500 -e 5000
If you find out that it seems stable after, e.g., 1500 ps, you can then use -b 1500 and analyse until the end of the simulation (5000 ps in your case currently, but might need to be more).