This paper presents an original method to separate the residential electric load into its major components. The method is explained in the particular case of space-heating, which is the most consuming electric end-use in France1. This is a source separation problem from a single mixture. The components to be retrieved are square signals characterized by a periodic regulation and a slowly timevarying duty cycles. A point process is used to model the electric load as a configuration of possibly overlapping square signals, given the priors on magnitude, duty cycle variations and the regulation periodicity. This stochastic process is simulated using a Reversible Jump Markov Chain Monte Carlo procedure. A simulated annealing scheme is used to achieve the posterior density maximization. First results on real data provided by Electricité de France are quite encouraging.