Arming at the current inaccurate problem of multi-feature power load prediction accuracy, in order to fully exploit the feature information such as time-series information and weather information in power load data, a variational mode decomposition (VMD)-long short-term memory (LSTM) neural network-light gradient boosting machine (LightGBM) based prediction model is proposed to optimize the problems of non-linearity, non-stationary and long memory of load data, and solve the problem of poor extraction of feature information for multi-feature prediction. The method first decomposes the characteristic mode components representing different scales with VMD, which reduces instability of the original sequence, while the residual amount of the decomposition represents the strongly nonlinear part of the load data, which is predicted by the characteristic strong algorithm. Each modal component is predicted by the single feature of LSTM, and then each component is added to the multi-feature using LightGBM for load prediction. Experimental comparison with current multi-characteristic power load forecasting models through example analysis is conducted, the proposed method MAE value is only 23% ~ 73% , and MAPE value can reach 0.37%, with better prediction accuracy.