To further improve the accuracy of short-term power load prediction, it is necessary to explore the hidden nonlinear relationships in load data at a deeper level. A short-term memory neural network based on quadratic modal decomposition of signal decomposition technology was proposed for short-term power load forecasting. Firstly, the proposed algorithm decomposed the original load data using the complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) algorithm. Secondly, the strong non-stationary components in the CEEMDAN decomposed components were decomposed by the variational mode decomposition (VMD) algorithm. at the same time, the number of VMD decomposition was optimized by using the central frequency method, and then the load subsequences obtained after two decompositions were fed into long short-term memory network (LSTM) for prediction, and the predicted results of the obtained components were overlaid. The results indicate that the method proposed in this article has significantly improved the accuracy of short-term power load forecasting results and model performance.