Recently，deep learning method has become the focus of academic research because of its achievement in areas such as image recognition. On this basis，this paper studies the feasibility of applying deep learning to the convergence criterion of power flow calculation，which is one of the fundamental calculation task in power systems. Compared with image recognition，the specificity of power flow calculation and the difficulty of applying deep learning in it are firstly discussed. After that，the network structure and training process of deep neural network（DNN） are presented as well as the method to solve the overfitting problem. Then according to the characters of power systems，feature vectors applied to the DNN model are constructed，including basic feature vectors，index set of reactive power adjustment capability and index set of power factor level. Finally，based on IEEE 14 system，1.6 million samples are generated and setting rules of hyper parameters are introduced，and calculation results show that high prediction accuracy can be achieved via the model and the feature vectors proposed in this paper.