For the sake of enhancing the power load forecasting accuracy, a two-stage short-term power load forecasting method based on VMDT-POA-DELM-GPR is proposed. In the first stage, the original load series is decomposed using Variational Mode Decomposition (VMD) to obtain the residual components after decomposition. Then, the time-varying filtering empirical mode decomposition (TVF-EMD) method is used for feature extraction; Next, a deep Extreme learning machine (DELM) model is established for all Subsequence, and Pelican optimization algorithm (POA) is used to optimize the parameters. The initial load prediction value is obtained by adding the prediction value of each Subsequence. In the second stage, the POA-DELM model is used to predict the error components; All Subsequence prediction values and error prediction values in the first stage are input into the Gaussian process Regression (GPR) model as features to obtain the final load prediction results. The results show that the Root-mean-square deviation (RMSE) and Mean absolute error (MAE) of the two-stage model are 4%~77% and 4%~76% of the comparison model respectively, while the average percentage error (MAPE) is only 0.0678%, which can effectively improve the accuracy of power load forecasting.