Energy-Aware Federated Learning in Satellite Constellations

TL;DR

This paper introduces an energy-aware scheduling method for federated learning in satellite constellations, significantly extending satellite battery life while maintaining model convergence.

Contribution

It presents a novel energy-aware computation scheduler that reduces battery consumption in satellite federated learning without affecting convergence speed.

Findings

Battery lifetime increased by over 3x with the new scheduler.

Energy management does not compromise model training convergence.

Scheduler effectively balances energy use and computational demands.

Abstract

Federated learning in satellite constellations, where the satellites collaboratively train a machine learning model, is a promising technology towards enabling globally connected intelligence and the integration of space networks into terrestrial mobile networks. The energy required for this computationally intensive task is provided either by solar panels or by an internal battery if the satellite is in Earth's shadow. Careful management of this battery and system's available energy resources is not only necessary for reliable satellite operation, but also to avoid premature battery aging. We propose a novel energy-aware computation time scheduler for satellite FL, which aims to minimize battery usage without any impact on the convergence speed. Numerical results indicate an increase of more than 3x in battery lifetime can be achieved over energy-agnostic task scheduling.