Load forecasting should fully consider nonstationary and nonlinear of the load. Local wave can decompose the load to get a series of stationary components, but the problem of end effects still exists. In order to reduce the nonlinear of load sequence to improve the accuracy of load forecasting, this paper introduces the four-midpoint estimation method to reduce the error on end effect based on the decomposition of local wave method. According to the characteristics of each component and least squares support vector machine, the paper uses optimal parameters to establish the DEMD-LSSVM prediction model for each component, then reconstructing to obtain the prediction results. Finally the paper takes actual load of power grid to make simulations on normal rest days and special holiday, compared with DEMD-SVM prediction model, to prove the validity and practicability of DEMD-LSSVM prediction model.