Computation-aware Energy-harvesting Federated Learning: Cyclic Scheduling with Selective Participation

TL;DR

This paper tackles energy efficiency in energy-harvesting federated learning by introducing FedBacys, a battery-aware cyclic participation framework that clusters clients by remaining energy and performs intra-group aggregation before inter-group multicast. A more energy-saving variant, FedBacys-Odd, implements selective (odd-chance) participation to further reduce energy consumption without hurting convergence. The authors provide a convergence analysis under standard smoothness and stochastic gradient assumptions, and derive bounds that incorporate the energy-driven participation probability $oldsymbol{ eta}$, with improvements under a Polyak–Łojasiewicz condition. Empirical evaluations on CIFAR-10 demonstrate superior energy efficiency and robustness compared to multiple baselines, with clear energy-accuracy trade-offs controlled by system parameters such as the epoch length $S$, local training cost $oldsymbol{ ext{kappa}}$, charging probability $oldsymbol{ extit{delta}}$, and battery capacity $E_{ ext{max}}$. Overall, the work offers a rigorous, scalable approach to sustainable EHFL through battery-aware scheduling and selective participation, with practical implications for energy-constrained edge learning.

Abstract

Federated Learning (FL) is a powerful paradigm for distributed learning, but its increasing complexity leads to significant energy consumption from client-side computations for training models. In particular, the challenge is critical in energy-harvesting FL (EHFL) systems where participation availability of each device oscillates due to limited energy. To address this, we propose FedBacys, a battery-aware EHFL framework using cyclic client participation based on users' battery levels. By clustering clients and scheduling them sequentially, FedBacys minimizes redundant computations, reduces system-wide energy usage, and improves learning stability. We also introduce FedBacys-Odd, a more energy-efficient variant that allows clients to participate selectively, further reducing energy costs without compromising performance. We provide a convergence analysis for our framework and demonstrate its superior energy efficiency and robustness compared to existing algorithms through numerical experiments.

Figures (7)

• Figure 1: A schematic view of energy-harvesting FL based on probabilistic decision rules.
• Figure 2: An example of a FedBacys network with $N=10$ users assigned into $G=3$ groups. Within each group, (1) each user transmits local updates to the group's hub if they satisfy the conditions to send local updates; (2) the hub aggregates the received updates; (3) the hub sends the updated intra-group model to the next group by multicast. The final group's hub sends the aggregated model to the server, which is sent to Group 1 of the new epoch. The hub users can be changed at the beginning of every epoch.
• Figure 3: Test accuracies from operating different FL algorithms with respect to epochs for $G=5$.
• Figure 4: Energy consumption of FedBacys and FedBacys-Odd with respect to local training energy requirement ($\kappa$) and the number of slots comprising one global communication round ($S$). ($\delta=0.5$ and $G=5$).
• Figure 5: Test accuracy of FedBacys and FedBacys-Odd with respect to $\kappa$ and $S$. ($\delta=0.5$ and $G=5$).

Theorems & Definitions (14)

• Theorem 1
• proof
• proof
• Theorem 2
• proof
• proof
• Lemma C.1
• proof
• Lemma C.2
• proof