Enhancing anomaly detection in distributed power systems using autoencoder-based federated learning.

Area of Science:

• Electrical Engineering
• Computer Science
• Artificial Intelligence

Background:

• The proliferation of Internet-of-Things (IoT) devices in electric power systems increases complexity and necessitates robust monitoring.
• Traditional anomaly detection methods face challenges with large datasets, high response times, and data leakage due to centralized processing.
• Deep learning and machine learning require substantial centralized training data, posing scalability and privacy concerns.

Purpose of the Study:

• To propose a novel Autoencoder-based Federated Learning (AEFL) method for accurate anomaly detection in distributed power systems.
• To address the limitations of centralized approaches, including high response times and data leakage.
• To develop a decentralized anomaly detection model that enhances efficiency and data privacy for IoT power systems.

Main Methods:

• Integration of Autoencoder (AE) and Federated Learning (FL) networks to create a hybrid FLAE model.
• Decentralized training of anomaly detection models directly on IoT devices.
• Utilizing AE for feature extraction and anomaly identification within the federated learning framework.

Main Results:

• The FLAE method achieves high accuracy in detecting anomalies in power consumption data.
• Decentralized training significantly reduces response time compared to centralized methods.
• The approach effectively mitigates data leakage issues by eliminating the need for data transfer.

Conclusions:

• The proposed FLAE method offers an effective and privacy-preserving solution for anomaly detection in IoT-enabled electric power systems.
• Decentralized learning is crucial for handling the scale and complexity of modern power grids.
• FLAE demonstrates the potential for real-time, secure anomaly detection without compromising sensitive data.