Transformer load has strong randomness. It is difficult to predict, but it can more accurately grasp the change of power system load. In the current intelligent algorithm, the prediction accuracy is low because of the small number of layers. In this paper, a deep belief network (CEEMD-DBN) model optimized by the complementary ensemble empirical mode decomposition (CEEMD) algorithm was proposed, and the evaluation standard for evaluating the accuracy of load forecasting was given. Firstly, the model uses the CEEMD algorithm to decompose the preprocessed and normalized transformer load sequence into different stationary sub-frequency sequences. Then, an independent sub-DBN model is established based on the decomposed sub-frequency sequences. The output of the sub-DBN model corresponds to the predicted value of the sub-frequency sequence. Finally, the predicted values of the sub-frequency sequences are combined again to obtain the estimated values of all transformer load. The reliability of the model is verified by the case.