In this paper we tackle a problem of one-year ahead energy demand estimation from macroeconomic variables. A modified Harmony Search (HS) algorithm is proposed to this end, as one of the novelties of the paper. The modifications on the proposed HS include a hybrid encoding, with a binary part to carry out a feature selection, and a real part, to select the parameter of a given prediction model. Some other adaptation focussed on the HS operators are also introduced. We study the performance of the proposed approach in a real problem of Energy demand estimation in Spain, from 14 macroeconomic variables with values for the last 30 years, including years of the crisis, from 2008. The performance of the proposed HS with feature selection is excellent, providing an accurate one year ahead prediction that improves previous proposals in the literature.