Abstract: Ensuring the integrity of the distribution network load data is the basis for data statistics and business analysis.A load forecasting algorithm based on deep belief networks was proposed to solve the problem of Guangzhou University Town load data missing caused by long-time disrepair of meters or wrong reading of meters.This algorithm estimates missing values to make sure the stable running of power system by the model of time series forecasting. The deep belief networks which is composed of a certain number of restricted Boltzmann machine obtains the initial value of the network model by unsupervised training algorithm.Finally,the prediction training model was obtained by top-down supervised learning.So as to improve the global search ability of the training model to avoid falling into a local optimal solution,the particle swarm optimization algorithm was used to optimize the model to obtain the global optimal solution. By comparing the prediction indexes of several prediction training models,the accuracy and effectiveness of the proposed prediction model were verified.