Towards Energy-Aware Federated Traffic Prediction for Cellular Networks

This project looks at the trade‑off between how well we can predict cellular‑network traffic and how much energy that prediction uses. 5G and future networks depend on deep‑learning models to allocate resources and spot problems. Those models need huge data sets, which raises privacy and bandwidth issues when the data are stored in one place. Federated learning (FL) solves the privacy problem by training the model on many base stations instead of a central server. A common oversight is that the energy used—and the carbon emitted—by all those distributed training jobs is rarely measured. To fix that, we built a method that tests several deep‑learning architectures inside a federated‑learning framework,

using real traffic data from base stations. Key parts of the method: – Run local training epochs on each base station. – Aggregate the locally trained models into a global model. – Calculate a new sustainability indicator that translates the computation into carbon‑emission numbers. By comparing model size, how fast it converges, and the resulting carbon cost, we create a clear way to pick architectures that save energy. The results show that the small accuracy improvements you get from larger models often come with a much larger carbon footprint. This helps engineers design communication systems that are both privacy‑preserving and environmentally friendly.

Message: Hi FE, I need help with “Towards Energy-Aware Federated Traffic Prediction for Cellular Networks” in “Environmental Engineering”