Arrears prediction for electricity customer through Wgan-Gp

It is not common for electric company that arrears can be found accurately and in time. This phenomenon leads to shortage of arrears data. It also cause process of predicting arrears can not be further developed. Stratified sampling has become a general solution to the problem. In this paper, a new method for arrears prediction is proposed, which uses Wgan-Gp (Improved Training of Wasserstein GANs) to simulate the real data and then generate fake data to predict arrears with DBN. Electric power data (256 indicators), like the image pixel (16*16), is used as input of the Wgan-Gp to train the generator. For the first time, this paper convert the prediction of arrears time to prediction of arrears interval. Next, we design a series of relevant indicators for experiment, which proves that the prediction accuracy can be improved with analogue data generated by Wgan-Gp.