Dynamic information network is a new challenging problem in the field of current complex networks. Research on network evolution contributes to analyzing the network structure, understanding the characteristics of the network, and finding hidden network evolution rules, which has important theoretical significance and application value. The study of the network structure evolution is of great importance in getting a comprehensive understanding of the behavior trend of complex systems. However, the network structure is difficult to represent and quantify. And the evolution of dynamic networks is temporal, complex, and changeable, which increases the difficulty in analysis. This study introduces "role" to quantify the structure of dynamic networks and proposes a role-based model, which provides a new idea for the evolution analysis and prediction of network structure. As for the model, two methods to explain the role are given. To predict the role distributions of dynamic network nodes in future time, this study transforms the problem of dynamic network structure prediction into role prediction, which can represent the structural feature. The model extracts properties from historical snapshots of sub-network as the training data and predicts the future role's distributions of dynamic network by using the vector autoregressive method. This study also proposes the method of dynamic network structure prediction based on latent roles (LR-DNSP). This method not only overcomes the drawback of existing methods based on transfer matrix while ignoring the time factor, but also takes into account of possible dependencies between multiple forecast targets. Experimental results show that the LR-DNSP outperforms existing methods in prediction accuracy.