Asynchronous federated learning with partial weights aggregation for energy consumption forecasting.

Area of Science:

• Energy Systems
• Machine Learning
• Cybersecurity

Background:

• Accurate energy forecasting is crucial for grid management and renewable energy integration.
• Smart meter data privacy is a significant concern due to potential exposure of sensitive user information.
• Federated Learning (FL) enables collaborative model training while preserving data privacy.

Purpose of the Study:

• To develop a privacy-preserving asynchronous federated learning framework for energy forecasting.
• To address limitations of synchronous FL, including training delays and communication overheads.
• To improve the efficiency and accuracy of energy consumption predictions.

Main Methods:

• Proposed an asynchronous federated learning framework for continuous global model updating.
• Introduced a federated asynchronous adaptive aggregation mechanism with dynamic learning rate adjustment.
• Implemented a partial aggregation strategy for Long Short-Term Memory (LSTM) models, exchanging only a subset of weights.

Main Results:

• The asynchronous adaptive strategy outperformed the classic FedAvg approach in energy forecasting.
• Maintained prediction accuracy comparable to personalized FedAvg while significantly reducing communication costs.
• Demonstrated statistically significant improvements over the classic FedAsync algorithm across various client groups.

Conclusions:

• The proposed asynchronous federated learning framework effectively enhances energy forecasting privacy and efficiency.
• Adaptive aggregation and partial model exchange are key to overcoming synchronous FL limitations.
• This approach offers a promising solution for secure and accurate energy consumption prediction in smart grids.