[Paper Review] Federated Learning for Short-term Residential Load Forecasting

Christopher Briggs, Fan Zhong+1 more|arXiv (Cornell University)|May 27, 2021

Privacy-Preserving Technologies in Data4 citations

TL;DR

This paper proposes a federated learning (FL) approach for short-term residential load forecasting that preserves consumer data privacy by training models locally on smart meters while aggregating global model updates. It demonstrates that FL combined with hierarchical clustering (FL+HC) and a personalization step achieves up to 5% better forecasting accuracy than local models, with approximately 10x less computation, making privacy-preserving, high-performance load forecasting feasible.

ABSTRACT

Load forecasting is an essential task performed within the energy industry to help balance supply with demand and maintain a stable load on the electricity grid. As supply transitions towards less reliable renewable energy generation, smart meters will prove a vital component to facilitate these forecasting tasks. However, smart meter adoption is low among privacy-conscious consumers that fear intrusion upon their fine-grained consumption data. In this work we propose and explore a federated learning (FL) based approach for training forecasting models in a distributed, collaborative manner whilst retaining the privacy of the underlying data. We compare two approaches: FL, and a clustered variant, FL+HC against a non-private, centralised learning approach and a fully private, localised learning approach. Within these approaches, we measure model performance using RMSE and computational efficiency. In addition, we suggest the FL strategies are followed by a personalisation step and show that model performance can be improved by doing so. We show that FL+HC followed by personalisation can achieve a $\sim$5\% improvement in model performance with a $\sim$10x reduction in computation compared to localised learning. Finally we provide advice on private aggregation of predictions for building a private end-to-end load forecasting application.

Motivation & Objective

• To address privacy concerns hindering smart meter adoption by enabling decentralized training of load forecasting models.
• To evaluate federated learning (FL) and a clustered variant (FL+HC) against centralized and localized learning benchmarks.
• To assess computational efficiency and model performance trade-offs in FL-based load forecasting.
• To investigate the impact of a personalization step on model accuracy after FL training.
• To provide guidance on private aggregation of predictions for end-to-end private forecasting systems.

Proposed method

• The study employs Long Short-Term Memory (LSTM) networks for time-series load forecasting across residential energy consumption data.
• Federated learning (FL) is used to train a global model by aggregating local model updates from clients (households) without sharing raw data.
• A clustered variant, FL+HC, groups similar clients using K-means clustering before training specialized models per cluster to handle non-IID data.
• After global FL training, a personalization step fine-tunes each client’s model using their local data to improve individual performance.
• Computational efficiency is measured in terms of total training samples (millions), comparing FL, FL+HC, centralized, and localized training.
• Private aggregation of predictions is discussed as a final step to enable end-to-end privacy-preserving forecasting.

Experimental results

Research questions

• RQ1 Can federated learning achieve comparable or better forecasting performance than centralized learning while preserving data privacy?
• RQ2 How does the FL+HC clustering approach improve performance on non-iid residential load data compared to standard FL?
• RQ3 What is the computational cost of FL and FL+HC relative to localized learning, and can savings be achieved?
• RQ4 Does applying a personalization step after FL training significantly improve model accuracy for individual households?
• RQ5 How can private aggregation of predictions be implemented to complete a fully private end-to-end forecasting pipeline?

Key findings

• FL+HC followed by personalization achieves a ~5% improvement in model performance (RMSE) compared to fully localized learning.
• The best-performing FL+HC model required ~10x fewer training samples than localized learning, reducing computational cost significantly.
• FL+HC reduced average computation by 12.7x and the best model by 13.1x compared to localized training.
• The personalization step added minimal computation, requiring only a few fine-tuning epochs after global FL training.
• FL+HC outperformed standard FL and matched or exceeded centralized learning in accuracy while maintaining data privacy.
• Private aggregation of predictions is feasible and recommended as a final step in building a complete privacy-preserving forecasting system.