The development of power system informationization and the increasing integration of distributed generators and electric vehicle to distribution network have increased the complexity of power consumption mode and put forward higher requirements for the accuracy and stability of load forecasting. A short-term load forecasting method based on deep belief network is proposed. The method includes the network construction

the layer-by-layer pre-training of the model parameters

the supervised fine-tuning

and the application of the model. In the pre-training process of the model parameters

the Gaussian-Bernoulli restricted Boltzmann machine(GB-RBM)is used as the first module for stacking the deep belief network to deal more effectively with the multi-type real-valued input data. And the partially supervised training algorithm combined by unsupervised training algorithm and supervised training algorithm is used for pre-training. The Levenberg-Marquardt(LM)optimization algorithm is used to fine-tune the parameters obtained by the pre-training phase

which can help to converge faster to the optimal solution. Finally

the actual load data are used for test and the experiments results show that the method proposed has higher prediction accuracy in the case of large training samples and complicated load factors.