Power load features nonlinearity, non-stationarity and variability, so traditional time series forecasting methods cannot fully capture the complex characteristics of data and their evolution laws. To improve the accuracy of power load forecasting, an improved hybrid model of VMD-EEMD-LSTM is proposed. The model decomposes historical power load data through variational mode decomposition (VMD) to obtain intrinsic mode components and residual terms at different frequencies. Then, the residual terms after VMD decomposition are second-order decomposed through Ensemble Empirical Mode Decomposition (EEMD). The intrinsic mode functions (IMFs) obtained from VMD and EEMD decompositions are processed as features and input into the long-term and short-term memory network (LSTM) model for forecasting, and the overall forecasting results can be obtained after combination. Comparative experiments are carried out with the actual power load data of a certain area in Jiangxi as the data set. The experimental results show that compared with the VMD-LSTM model, the root mean square error (RMSE) and mean absolute error (MAE) of the prediction results of the VMD-EEMD-LSTM model are reduced by 56.57% and 56.03% respectively. Compared with the EEMD-LSTM model, the RMSE and MAE of the prediction results are reduced by 10.06% and 9.01% respectively, verifying that the proposed model has higher prediction accuracy.