Chenhan Zhai

Cellular traffic prediction is essential for the energy-efficient and intelligent management of base stations (BSs). In this context, Federated Learning allows BSs to collaboratively train a shared model without transmitting raw data, thus significantly reducing communication overhead and enhancing privacy protection. Despite these advantages, most existing methods struggle to alleviate the coupled spatio-temporal heterogeneity of cellular traffic data across BSs, resulting in suboptimal global prediction performance. The main reason is that they decouple spatio-temporal relationships by modeling temporal and spatial properties independently, with limited consideration of their dependencies in cellular traffic. To overcome these limitations, we propose FedJST, a joint spatio-temporal federated cellular traffic prediction method that effectively integrates the temporal and spatial properties of cellular traffic from different BSs in two key aspects. Firstly, to capture spatio-temporal dependencies, we propose a multi-scale feature guidance strategy. This strategy leverages global multi-scale feature to incorporate shared spatio-temporal information, encouraging local models to capture complex spatio-temporal interactions. Secondly, we propose a spatio-temporal property-aware BS selection strategy that leverages a lightweight augmented dataset to prioritize representative BSs for training, enabling effective learning of diverse spatio-temporal patterns. We conduct extensive experiments across six cases on two real-world datasets, demonstrating that FedJST achieves average improvements of 5.95%, 4.43%, and 2.99% in RMSE, MAE, and R² metrics, respectively.