Abstract: Along with the gradual grid-connection of distributed renewable energy at the user side, relative to original actual load curve the form of users’ net load curve monitored by electric meter becomes more unstable, so that the predicted accuracy of users’ net load is extremely decreased. For this reason, based on the wavelet packet decomposition (abbr. WPD) and least squares support vector machine (abbr. LSSVM) a method to predict user side net load was proposed. Through the WPD of users’ net load time series data the high frequency components and the low frequency trend with more evident signal features were obtained, and the high-frequency detail components, which evidently fluctuated and contained many noise signals, were screened and rejected. Meanwhile, considering meteorological factors and such characteristics of LSSVM as high training efficiency and strong generalization ability, the trained LSSVM model was used to respectively predict and reconstruct other low-frequency components containing more effective load data information, and then superposed them to obtain final predicted value of net load. Results of instance analysis on users’ actual net load data of a certain region with high penetration of renewable energy show that under such a physical scene a higher load prediction accuracy can be obtained than by traditional prediction method.