You could store the standard deviation as well as the mean (or median) of the ensemble. That would maximally double your data volume (but would add quite a bit less with compression). The UK Met Office have done same nice work into how to optimally combine the severity and risk of an event into an impact measure