Power distribution network failure events will seriously affect the normal operation of social and economic life. Therefore, an effective method for predicting the power outage of a distribution network is necessary. The artificial intelligence method is used to analyze the power outage data of the distribution network. It is found that there are fewer times of power outages and the uneven distribution of the causes of the power outages in the distribution network. In order to predict the power outage situation of a distribution network in time and accurately, this paper proposes a modified power outage prediction model from the aspects of data set quality and overfitting prevention. First, it designs a cluster-based generative adversarial neural network to enhance the quality of the data set, then constructs a power outage prediction model based on a random cost-sensitive convolutional neural network (RandomCost-CNN). The RandomCost-CNN prediction algorithm adopts the idea of random sampling with replacement to design a random selection strategy of the loss functions, so that it can alleviate the minority class (power outage class) overfitting and avoid a large number of false positives for the majority class (normal class). It can ensure that minority classes have better recall & accuracy and improve the generalization performance of the model. Experiments show that the proposed method can effectively predict the probability of a power outage event in the distribution network. It can play a good early warning role in the operation and maintenance management of the distribution network.