To address the challenges of short-term power load forecasting, a hybrid model, DWT-MFO-GRU, based on the combination of discrete wavelet transform (DWT), mothballing optimization algorithm (MFO) and gated recurrent unit (GRU) is proposed. The model firstly uses wavelet transform to decompose the load sequence into high-frequency components and The model firstly uses wavelet transform to decompose the load sequence into high-frequency components and low-frequency components, which are input into GRU with different initial weights and thresholds for prediction, and at the same time, MFO is used to optimize the initial weights and thresholds of GRU to obtain the optimal initial weights and thresholds. Secondly, they are assigned to the GRU to obtain the load prediction results of the high-frequency component and the low-frequency component, and finally both of them are reconstructed using the inverse wavelet transform (IWT) to obtain the final load prediction results. The experimental results show that compared with the traditional GRU, MFO-GRU and DWT-GRU models, the DWT-MFO-GRU significantly improves the key indexes such as the root mean square error, the mean absolute error and the mean absolute percentage error, which fully verifies that the synergistic framework of “Signal Preprocessing-Parameter Optimization” can be applied to the non-stationary load prediction. It fully verifies the effectiveness of the, signal preprocessing-parameter optimization” synergistic framework in non-stationary time-series load forecasting. The method provides a new idea for improving the decision support of power system dispatch and management, and has good prospects for popularization and application.