Aiming at the problem of unstable power load sequence, strong randomness, poor fitting effect and low prediction accuracy caused by direct input model, a fusion forecasting method of ensemble empirical mode decomposition (CEEMD) adding sets of complementary white noise and gated recurrent unit neural network (GRU) is proposed in this paper. For the empirical mode decomposition (EMD) to deal with the sequence of large interference signals, there is a modal aliasing problem, the method of CEEMD is proposed; and then, adding complementary white noise, decomposing the original sequence into sub-sequences of different scales and inputting GRU neural network, optimizing the network super parameters, and finally, the predicted results are obtained. The experiment results show that the reconstruction error is small and the prediction effect is good.