Resumen
Wireless traffic prediction is critical to the intelligent operation of cellular networks, such as load balancing, congestion control, value-added service promotion, etc. However, the BTS data in each region has certain differences and privacy, and centralized prediction needs to transmit a large amount of traffic data, which will not only cause bandwidth consumption, but may also cause privacy leakage. Federated learning is a kind of distributed learning method with multi-client joint training and no sharing between clients. Based on existing related research, this paper proposes a gradient similarity-based federated aggregation algorithm for wireless traffic prediction (Gradient Similarity-based Federated Aggregation for Wireless Traffic Prediction) (FedGSA). First of all, this method uses a global sharing enhanced data strategy to overcome the data heterogeneity challenge of multi-client collaborative training in federated learning. Secondly, the sliding window scheme is used to construct the dual channel training data to improve the feature learning ability of the model; In addition, to improve the generalization ability of the final global model, a two-layer aggregation scheme based on gradient similarity is proposed. The personalized model is generated by comparing the gradient similarity of each client model, and the central server aggregates the personalized model to finally generate the global model. Finally, the FedGSA algorithm is applied to wireless network traffic prediction. Experiments are conducted on two real traffic datasets. Compared with the mainstream Federated Averaging (FedAvg) algorithm, FedGSA performs better on both datasets and obtains better prediction results on the premise of ensuring the privacy of client traffic data.