An important direction of mobile communication technology is mobile traffic prediction. Reasonable and accurate mobile traffic prediction can not only effectively reduce network congestion and improve user service quality, but also help to allocate network resources reasonably and improve communication system efficiency. However, existing mobile traffic prediction methods have many drawbacks, such as the huge demand for single machine computing power, high channel resource occupancy, poor privacy protection, and low willingness to share data across sources and domains, which cannot meet the needs of mobile traffic prediction in intelligent mobile communication processes. This article proposes a method that combines cross source and cross domain data to accurately model and predict mobile traffic. This method describes how to use federated proximal techniques to achieve accurate mobile traffic prediction. We have conducted in-depth research and improvements in mobile traffic prediction, cross source and cross domain data collaborative training, and joint learning global model training. The proposed method was evaluated by comparing two widely used clustering methods. Experimental data shows that the proposed solution can improve the accuracy of mobile traffic prediction.