To explore the hidden information in power load data and improve the accuracy of short-term load forecasting, a hybrid prediction model based on variational mode decomposition (VMD), long-term and short-term memory network (LSTM), improved particle swarm optimization (IPSO) algorithm and gated recurrent unit neural network (GRU) was proposed with consideration of strong non-linearity and non-stationarity in power load. First, correlation analysis was employed to determine input factors. The VMD algorithm combined with sample entropy was employed to decompose the load data into a series of intrinsic mode function (IMF) and residual components, which then allowed for the rational determination of the decomposition levels and penalty factors. Then, these quantities were divided into low-frequency and high-frequency components based on the zero-crossing rate. The low-frequency components were forecasted using an LSTM network, while the high-frequency components were predicted with an IPSO-GRU network. Finally, the predicted results were reconstructed to obtain the final result of power load. Simulation results show that the proposed hybrid prediction model can effectively extract modal features and possesses higher predictive accuracy compared with alternative models.