In order to improve the short-term power load forecasting performance, a novel combined forecasting method based on pole symmetrical empirical mode decomposition (ESMD)-permutation entropy (PE) and adaptive deep belief network (ADBN) is proposed in this paper. In order to improve the prediction accuracy and reduce the complexity of the original load sequence and simplify the input of the prediction model, the ESMD method is first used to decompose the original load sequence into a series of modal functions of different complexity, and then, the permutation entropy is used to calculate the entropy value of each modal function and reconstruct modalities of similar complexity to obtain new subsequences; on the basis of comprehensive consideration of various influencing factors, this paper constructs different DBN prediction models for the new sequence, and finally, superimposes the prediction results; due to the unsupervised training stage in the DBN model, the learning rate usually adopts globally uniform constant parameters. The adaptive learning rate is introduced into the contrast difference (CD) algorithm. The convergence of the model is improved by automatically adjusting the learning rate, and the prediction accuracy is also improved. Through example analysis, the MAPE and RMSE values of the ESMD-PE-ADBN model proposed in the paper are 1.03% and 90.91 MW, respectively, and the prediction effect is the best.