Abstract: Prediction of daily peak load for next month is an important type of Medium Term Load Forecast (MTLF) for electrical power systems. In this study, a new MTLF strategy is proposed, which is composed of a new two stage feature selection technique and hybrid forecast engine. new nonlinear feature selection algorithm composed of Modified Relief (MR) and Mutual Information (MI) techniques is proposed for the MTLF and proposed forecast engine has MLP structure and a new hybrid training mechanism composed of LM learning algorithm and Differential Evolution (DE). Two test cases are considered to evaluate the efficiency of the proposed MTLF strategy: EUNITE competition data and Iran power system.