We consider the sequential short-term forecast of electricity consumption based on ensemble methods. That is, we use several possibly independent base forecasters and design meta-forecasters which combine the base predictions that are output by them. These meta-forecasters are extracted from the literature of sequential learning, namely, from the field called prediction with expert advice, and come with strong theoretical guarantees. The forecasters considered here may be specialized and need not output a prediction at all time indexes while the meta-forecasters have to. We first motivate review and adapt existing meta-forecasters to our setting and then describe the improvements obtained by these techniques on two data sets, a Slovakian one and a French one, respectively concerned with hourly and half-hourly predictions. These improvements lie in a reduced mean squared error but also in a more robust behavior with respect to large occasional errors.