Accurate prediction of wireless communication network traffic can assist operators in precise operation, improve communication network management, and reduce energy consumption. However, due to the highly complicated spatio-temporal dependence and the influence of multi-source cross domain data, the accurate prediction of cellular traffic is facing great challenges. In this work, we propose a Dense-convolutional-neural-network-based traffic prediction model for fusion of Multi-Source Data(MS-DCN). The model includes spatio-temporal module and external feature module. We leverage DenseUnit architecture to capture temporal characteristics with different degree of dependence and study spatial characteristics. In external feature module, the same DenseUnit architecture is employed to capture multi-soure factors. Spatiotemporal features and external features are effectively integrated to achieve accurate prediction of large-scale wireless communication traffic. In the experimental part, MS-DCN is proved to have higher prediction accuracy than the existing models on the actual cellular data set.