Deep Federated Adaptation: An Adaptative Residential Load Forecasting Approach with Federated Learning

Area of Science:

• Electrical Engineering
• Computer Science
• Artificial Intelligence

Background:

• Residential short-term load forecasting (STLF) is crucial for power system operations.
• Data-driven models require substantial data, but privacy concerns and limited data availability hinder model development.
• Non-identical and independent distribution (non-IID) data across households complicates accurate, personalized STLF.

Purpose of the Study:

• To address data limitations and privacy concerns in residential STLF.
• To develop a privacy-preserving model that overcomes non-IID data challenges.
• To improve the accuracy and computational efficiency of STLF for individual residences.

Main Methods:

• Proposed Deep Federated Adaptation (DFA), a federated transfer learning approach for STLF.
• Utilized federated learning to train a global model without compromising data privacy.
• Employed multiple kernel variant of maximum mean discrepancies (MK-MMD) for fine-tuning the global model to specific house data.

Main Results:

• DFA demonstrated superior forecasting performance compared to baseline models on real residential datasets.
• The federated architecture of DFA showed significant advantages in reducing computation time.
• Extended DFA framework with alternative transfer learning methods, all achieving good STLF performance.

Conclusions:

• Federated transfer learning, specifically DFA, offers an effective solution for privacy-preserving STLF.
• DFA successfully tackles challenges posed by limited and non-IID data in residential load forecasting.
• The proposed approach enhances prediction accuracy and computational efficiency for STLF applications.