GreenFlag - A Green Agentic Approach for Energy-Efficient Federated Learning

This rollup-banner showcases the paper "GreenFLag: A Green Agentic Approach for Energy-Efficient Federated Learning". It presents the system description of the paper, a visualization tool used for the simulation of the network and the work's evaluation results. This work was done during the EXIGENCE Project. As 6G networks shift toward distributed intelligence, Federated Learning (FL) enables private, on-device AI training but suffers from high carbon emissions. To solve this, NKUA and HWDU developed GreenFLag, an agentic resource orchestration framework that minimizes grid power reliance. Powered by Soft Actor-Critic (SAC) reinforcement learning, GreenFLag dynamically optimizes computational and communication resources, while taking into account available renewable resources. Tested on real-world Copernicus meteorological data, it cut grid energy consumption by an average of 94.8% compared to three state-of-the-art baselines - proving that future 6G edge AI can run almost entirely on green power without sacrificing performance