As wireless networks evolve toward Beyond-5G and 6G, distributed Artificial Intelligence (AI) is becoming a core enabler of automation, adaptability, and real-time decision-making. Federated Learning (FL) allows multiple devices to collaboratively train AI models without sharing raw data, enhancing privacy and scalability.
However, this shift comes with a significant challenge: energy consumption. Distributed training across edge devices introduces substantial computational and communication overhead, increasing reliance on grid power and raising sustainability concerns.
Within the EXIGENCE project, we address this challenge by developing energy-aware AI solutions that align high performance with environmental responsibility. One such contribution is GreenFLag, a novel framework designed to make FL significantly more energy-efficient.
The Challenge: Energy-Aware Federated Learning
While FL reduces data transfer and improves privacy, it introduces new complexities in resource management. Each participating device must perform local computation and communicate model updates, both of which consume energy. In real-world deployments, devices may also have access to renewable energy sources, such as solar or wind. However, these sources are inherently variable, making it difficult to optimally balance:
- Computational workload
- Communication resources (bandwidth, transmission power)
- Available renewable energy
- Grid energy usage
Traditional approaches to FL orchestration do not account for this dynamic energy landscape, often leading to inefficient resource allocation and unnecessary reliance on grid power.
The Solution: GreenFLag Framework
GreenFLag introduces an agentic, Reinforcement Learning(RL)-based resource orchestration framework that dynamically optimises FL processes with energy efficiency as a primary objective.
At its core, GreenFLag:
- Prioritises renewable energy usage, treating grid power as a fallback
- Jointly optimises computation, transmission power, and bandwidth allocation
- Adapts decisions over time using a Soft Actor-Critic (SAC) RL agent
- Ensures system feasibility through constraint-aware scheduling and penalty mechanisms
The framework operates across iterative FL rounds, learning how to allocate resources in a way that minimises long-term grid energy consumption while maintaining model performance.
Another key component is a bandwidth scheduler, which ensures realistic communication constraints by managing channel contention and preventing resource over-allocation.
Key Innovation: Renewable-Aware Intelligence
A distinguishing feature of GreenFLag is its integration of renewable energy into the decision-making loop.
Instead of treating all energy sources equally, the system follows a hierarchical strategy:
- Use locally harvested renewable energy
- Use stored energy (battery)
- Resort to grid energy only when necessary
This design allows the system to adapt to fluctuating environmental conditions (e.g., solar radiation, wind availability), enabling energy-aware AI orchestration at scale.
Performance and Impact
GreenFLag has been evaluated using realistic wireless network simulations and real-world renewable energy data from the Copernicus dataset.
The results demonstrate:
- Up to 94.8% reduction in grid energy consumption compared to state-of-the-art approaches
- Significant reduction in overall energy usage
- Comparable model accuracy and convergence speed, ensuring no performance trade-offs
- Strong robustness under varying conditions, including limited or unavailable renewable energy
These findings confirm that sustainability can be achieved without compromising the effectiveness of distributed AI systems.
Contribution to EXIGENCE Vision
GreenFLag directly supports the EXIGENCE vision of energy-efficient AI-native wireless networks. By embedding sustainability into the core of AI orchestration, it demonstrates how intelligent systems can:
- Reduce environmental impact
- Improve operational efficiency
- Adapt dynamically to real-world constraints
This work highlights the importance of cross-layer optimisation, where AI, networking, and energy systems are jointly considered to achieve meaningful gains.
Conclusion
As AI continues to transform wireless networks, ensuring its sustainability becomes critical. GreenFLag represents a step toward green FL, where intelligent orchestration aligns performance with environmental responsibility.
By leveraging RL and renewable-aware strategies, the framework demonstrates that future networks can be both powerful and energy-efficient, supporting Europe’s broader goals for sustainable digital infrastructure.
References
Theodora Panagea et al. GreenFLag: A Green Agentic Approach for Energy-Efficient Federated Learning.2026. arXiv: 2603.29933 [cs.NI]. url: https://arxiv.org/abs/2603.29933.
Authors
This work is a collaborative effort by researchers from the National and Kapodistrian University of Athens and Huawei Munich Research Center, contributing to the broader EXIGENCE research ecosystem.
SCANLAB
Theodora Panagea is currently pursuing her Masters Degree at the Department of Informatics and Telecommunications, National and Kapodistrian University of Athens (NKUA). She received her Bachelor of Science (B.Sc.) from the same department, with a specialization in “Computer Science Foundations”. She is currently working as a Research Associate at the Software Centric and Autonomic Networking (SCAN) Laboratory at NKUA. Her main topics of interest are machine learning, artificial intelligence, algorithms, and cognitive science.
SCANLAB
Lina Magoula is currently a Senior PhD candidate and is working as Technical Manager at the SCAN Laboratory of National and Kapodistrian University of Athens, performing active research in applied AI for next generation networks. Since 2021, she has been working as Technical Manager of the SCAN laboratory, actively involved in the design and development of AI-empowered solutions for intelligent resource orchestration for 5G+ networks. Her main areas of interest are Machine Learning, Artificial Intelligence, 5G and Beyond Networks, Distributed Learning, Green Communications, Smart Resource Orchestration and Network Function Virtualization.
SCANLAB
Nikolaos Koursioumpas is a Senior PhD Candidate and works as Technical Project Manager at the SCAN Laboratory of National and Kapodistrian University of Athens. He received his B.Sc and M.Sc from the Department of Informatics and Telecommunications with specialisation in “Computer Networking and Telecommunications”. He is distinguished for his strong background in the design and development of distributed AI/ML algorithms for dynamic resource allocation through Context and Energy-Aware Resource Management. His main areas of interest include: Distributed AI, Green 5G+ and 6G Networks
Huawei Technology Dusseldorf GmbH
Ramin Khalili is a principal researcher with over 20 years of pioneering innovation in resource management for distributed systems and wireless networks, and more than 7 years of advanced work in efficient AI, multi-agent reinforcement learning (MARL) and Agentic AI, and sustainable computing.