Operational model is the "real" model--the set of equations and parameterizations that are the "best" (heavy emphasis on quotes there). To sort of test the stability of an operational run, several other versions are run simultaneously with small differences in the starting conditions. These are the ensembles.
If the operational model and ensembles agree, you tend to believe that the operational model has a good grasp on the setup--so good that small perturbations to the model won't change the overall picture too much. If there is either 1) disagreement between the operational and the ensembles or 2) huge spread in the ensembles, that signals more uncertainty in the operational output.
Here's one article I found in a quick google search. Web is loaded with them.
http://www.emc.ncep.noaa.gov/gmb/ens/in ... etbak.html