Abstract: The integrity and validity of load data are of great significance in load forecasting and other applications. The traditional completion methods for missing data lack the mining of the power load and the various external spatio-temporal correlation information, so it is difficult to obtain high-precision completion results. In this paper, a double-channel generative adversarial network is proposed to complete the missing load data. First, a third-order load tensor is constructed according to the periodic variation characteristics of load and the spatio-temporal correlation, and a variety of external factors affecting load changes are constructed as the third-order auxiliary information tensors. Then, in order to meet the double input requirements of the two tensors, a double-channel mechanism is introduced at the input layer of the generative adversarial network, and the features of the tensors are extracted by convolution and deconvolution operations. In order to improve the training effect and completion accuracy of network on tensor data, the tensor decomposition loss is introduced into the original loss function, and an improved chaotic particle swarm optimization algorithm is used to jointly optimize the hyperparameters and the network. Finally, the data completion experiment is carried out on the real load data set. The results show that the proposed method can accurately complete the load data with random loss rate of less than 50% and continuous loss of less than 3 days.