This work considers the problem of actively managing the power consumption of a large number of thermostically controlled loads (TCLs), namely a TCL population, and a case-study of household refrigerators. Control is performed using a new randomized actuation that consists of switching units on and off at given rates, while at the same time respecting the nominal constraints on each individual unit. Both the free and the controlled behavior of individual TCLs can be aggregated, making it possible to handle a TCL population as if it were a single system. The aggregation method uses the distribution of the TCLs individual states across the population. The distribution approach has two main advantages. It scales excellently since the computational requirements do not increase with the number of units, and it allows data from individual units to be used anonymously, which solves privacy concerns relevant for consumer adoption.