Researchers Eunjeong Jeong, Giovanni Perin, Howard H. Yang, and Nikolaos Pappas, affiliated with Chalmers University of Technology in Sweden, have developed a novel approach to improve the efficiency of federated learning (FL) in energy-harvesting scenarios. Their work, published in the journal IEEE Transactions on Mobile Computing, addresses a significant challenge in the energy industry’s growing interest in edge computing and machine learning applications.
Federated learning allows multiple devices to collaboratively train a machine learning model without sharing their raw data. This approach is particularly useful in the energy sector, where edge devices like smart meters, sensors, and IoT devices generate vast amounts of data that can be used to optimize energy distribution, predict demand, and improve overall grid efficiency. However, training deep neural networks (DNNs) on these resource-constrained devices can be energy-intensive, often consuming more energy than the communication processes involved in FL.
The researchers identified that existing energy-harvesting federated learning (EHFL) strategies often waste energy due to redundant local computations. To tackle this issue, they proposed a lightweight client scheduling framework that uses a semantics-aware metric called Version Age of Information (VAoI). VAoI quantifies the timeliness and significance of local updates, helping to predict which devices will contribute the most valuable updates to the global model.
One of the key innovations in their approach is the introduction of a feature-based proxy that estimates model redundancy using intermediate-layer extraction from a single forward pass. This dramatically reduces the computational complexity typically associated with VAoI, making it practical for use in resource-constrained edge devices.
The researchers conducted experiments under extreme non-independent and identically distributed (non-IID) data distributions and scarce energy availability, demonstrating that their framework achieves superior learning performance while reducing energy consumption compared to existing baseline selection policies.
For the energy industry, this research offers a practical solution to optimize the use of energy-harvesting edge devices in federated learning scenarios. By reducing redundant computations and improving the efficiency of local updates, the proposed framework can help lower the energy footprint of machine learning applications in the energy sector. This can lead to more sustainable and cost-effective implementations of smart grid technologies, demand response systems, and other energy management applications that rely on edge computing and machine learning.
Source: Jeong, E., Perin, G., Yang, H. H., & Pappas, N. (2023). Feature-Based Semantics-Aware Scheduling for Energy-Harvesting Federated Learning. IEEE Transactions on Mobile Computing.
This article is based on research available at arXiv.