October 1, 2024

Addressing Heterogeneity in Federated Load Forecasting via Personalization Layers

SBShourya BoseUniversity of California, Santa Cruz YZYu ZhangTongji University KKKibaek KimArgonne National Laboratory

Key Points

Personalization layers significantly improve load forecasting performance compared to basic federated learning approaches.
In extensive simulations on three datasets, models showed reduced communication bandwidth usage while enhancing accuracy.
Federated learning enables privacy-preserving training of load forecasting models, which tackles data privacy concerns in energy usage.
Addressing heterogeneity in client data, the proposed PL-FL framework demonstrates clear advantages over traditional methods.

Abstract

The advent of smart meters has enabled the pervasive collection of energy consumption data for training short-term load forecasting models. Addressing privacy apprehensions, federated learning (FL) has emerged as a privacyconscious methodology for model training. Nevertheless, the efficacy of trained models diminishes as client data exhibits heterogeneity. In this study, we introduce personalization layers within a comprehensive framework termed PL-FL, aimed at enhancing load forecasting. PL-FL demonstrates superior performance compared to FL and purely local training methods, while also demanding lower communication bandwidth than FL. This advancement is validated through extensive simulations conducted on three datasets sourced from the NREL ComStock repository.

Demander à l'IA

Bookmark

View Full Paper