[Paper Review] Secure Federated Learning for Residential Short Term Load Forecasting

Joaquín Delgado Fernández, Sergio Potenciano Menci+2 more|arXiv (Cornell University)|Nov 17, 2021

Smart Grid Energy Management68 references4 citations

TL;DR

This paper proposes a secure federated learning framework for residential short-term load forecasting that preserves data privacy through differential privacy and secure aggregation. It achieves near-perfect privacy budgets (1.39, 10⁻⁵) and (2.01, 10⁻⁵) with minimal impact on forecasting accuracy, enabling decentralized, collaborative learning without sharing raw smart meter data.

ABSTRACT

The inclusion of intermittent and renewable energy sources has increased the importance of demand forecasting in power systems. Smart meters can play a critical role in demand forecasting due to the measurement granularity they provide. Consumers' privacy concerns, reluctance of utilities and vendors to share data with competitors or third parties, and regulatory constraints are some constraints smart meter forecasting faces. This paper examines a collaborative machine learning method for short-term demand forecasting using smart meter data as a solution to the previous constraints. Privacy preserving techniques and federated learning enable to ensure consumers' confidentiality concerning both, their data, the models generated using it (Differential Privacy), and the communication mean (Secure Aggregation). The methods evaluated take into account several scenarios that explore how traditional centralized approaches could be projected in the direction of a decentralized, collaborative and private system. The results obtained over the evaluations provided almost perfect privacy budgets (1.39,$10e^{-5}$) and (2.01,$10e^{-5}$) with a negligible performance compromise.

Motivation & Objective

• Address privacy concerns in residential smart meter data sharing for power system forecasting.
• Overcome regulatory and commercial barriers to data sharing among utilities and vendors.
• Develop a decentralized, collaborative machine learning approach that maintains data confidentiality.
• Ensure model privacy through differential privacy and communication security via secure aggregation.
• Evaluate the trade-off between privacy guarantees and forecasting performance in realistic scenarios.

Proposed method

• Employ federated learning to train load forecasting models locally on residential smart meter data without centralizing raw data.
• Apply differential privacy to the model updates to protect individual data privacy, with carefully tuned noise injection.
• Use secure aggregation to cryptographically combine model updates from multiple clients, preventing exposure of individual contributions.
• Design a decentralized training pipeline where clients train locally and only share encrypted model updates with a central parameter server.
• Evaluate the system across multiple realistic scenarios to assess privacy-accuracy trade-offs.
• Calibrate privacy budgets (ε) using theoretical bounds to ensure strong privacy guarantees while maintaining model utility.

Experimental results

Research questions

• RQ1 How can residential short-term load forecasting be achieved in a decentralized manner without sharing raw smart meter data?
• RQ2 To what extent can differential privacy and secure aggregation preserve individual privacy while maintaining forecasting accuracy?
• RQ3 What are the achievable privacy budgets (ε) in a federated learning setup for residential load forecasting?
• RQ4 How does the performance of the federated model compare to centralized baselines under equivalent privacy constraints?
• RQ5 What is the trade-off between privacy guarantees and forecasting accuracy in real-world residential load forecasting scenarios?

Key findings

• The proposed federated learning framework achieves a privacy budget of 1.39 with a privacy loss of 10⁻⁵, indicating strong privacy protection.
• A second scenario achieves a privacy budget of 2.01 with the same privacy loss of 10⁻⁵, demonstrating robustness across configurations.
• The privacy-preserving mechanisms result in negligible performance degradation compared to non-private centralized models.
• Secure aggregation successfully prevents inference of individual client data from model updates during aggregation.
• Differential privacy effectively limits the risk of membership inference attacks on individual households.
• The system enables collaborative forecasting without data sharing, satisfying regulatory and commercial constraints.