Research on Federated Learning Traffic Prediction Algorithm Based on Deep Autoregressive Networks

Wireless traffic prediction is critical for cellular networks to achieve intelligent network operation, and it is well known that the adoption of traffic prediction can facilitate resource allocation, improve efficiency, and ultimately enable intelligent cellular networks. Existing prediction methods usually use centralized training architectures and require the transmission of large amounts of Cellular traffic data, which may cause latency and privacy issues in some scenarios. In this paper, we propose a new efficient prediction framework for wireless network traffic, called Federated Learning Traffic Prediction Algorithm based on Deep Autoregressive Networks (FL-DeepAR), by which multiple edge clients can collaborate to train high-quality prediction models independently. The framework uses a homogenized global shared augmented dataset strategy to overcome the challenge of statistical heterogeneity of data; a deep autoregressive network is used to capture the multidimensional data features that will have an impact on the future values of the base station network traffic; a gradient-based similarity aggregation scheme is proposed to aggregate the global model by comparing the gradient similarity of the models across clients. The experimental results show that FL-DeepAR outperforms the mainstream FedAvg and FedAtt methods, and the proposed method also improves the prediction performance on both datasets compared with the traditional training method and the non-federal learning training method.