Abstract: Abstract: Wireless communication network traffic prediction is of great significance to operators in network construction, base station wireless resource management and user experience improvement. However, the existing centralized algorithm model faces the problems of complexity and timeliness. Therefore, it is difficult to meet the traffic prediction of the whole city scale. A distributed wireless communication traffic prediction framework was proposed to realize traffic prediction based on single grid base station with low complexity and communication overhead. Based on the distributed architecture, a wireless traffic prediction model based on federated learning was proposed. Each grid traffic prediction model was trained synchronously. JS (Jensen-Shannon, JS) divergence was used to select grid flow models with similar flow distributions, and federated average algorithm was used to fuse the parameters of the grid flow models with similar flow distributions, so as to accurately describe the regional traffic. In addition, the flow in different areas within the city is highly differentiated. On the basis of the algorithm, a federation construction method based on a coalitional game was proposed. Combined with super-addivity, the grids were used as participants in the coalitional game, and screened. The core of the coalitional game and the Shapley value were introduced for profit distribution to ensure the stability of the alliance, thereby improving the accuracy of model prediction. Experiments show that compared with grid-independent training, the prediction error of the model decreases most significantly in the suburbs, by 28.7%, at most by 7.8% in the urban area, and at most by 4.7% in the downtown area. Compared with the grid-centralized training, the prediction error of the three regions decreases by 43.8% to 79.1%.

Key words: Federated learning, Cloud-Edge collaboration, Wireless traffic prediction, JS divergence, Coalitional game

摘要： 摘 要: 无线通信网络流量预测对运营商进行网络建设、基站无线资源管理以及用户体验提升具有重要意义。然而，现有的集中式算法模型面临着复杂性和时效性问题，因此，难以满足城市全域尺度的通信流量预测需求。提出一份基于分布式的云边协同无线通信流量预测框架，能够以较低的复杂度和通信开销实现基于单栅格基站的流量预测。在分布式架构基础上，提出了基于联邦学习的无线通信流量预测模型，各个栅格流量预测模型同步训练，中心云服务器利用JS（Jensen-Shannon）散度挑选出流量分布相似的栅格流量模型，并采用联邦平均算法对具有相似流量分布的栅格流量模型的参数进行融合，从而达到既能够提高模型的泛化性同时又能够保持对本地流量精准刻画的目的。此外，城市范围内不同地区流量具有高度差异化的特征，在算法的基础上，提出了基于合作博弈的联邦训练方法，将栅格作为合作博弈的参与者，结合超可加性准则对栅格进行筛选，并引入合作博弈的核以及夏普利值进行收益分配以确保联盟的稳定性，从而提高模型预测的准确性。实验表明，以短消息业务流量为例，与栅格独立式训练相比，模型预测误差下降在郊区最为明显，下降了28.7%，在市区至多提升了7.8%，在市中心至多提升了4.7%。与栅格集中式训练相比，三个区域的模型预测误差下降范围在43.8%到79.1%。

关键词: 联邦学习, 云边协同, 无线流量预测, JS散度, 合作博弈