FedBranched: Leveraging Federated Learning for Anomaly-Aware Load Forecasting in Energy Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) is gaining traction for edge computation and privacy.
• A key challenge in FL is managing high data diversity across clients.
• Diverse datasets can lead to anomalous local model behavior and reduced accuracy.

Purpose of the Study:

• To introduce FedBranched, a novel clustering-based framework for federated learning.
• To address the challenge of client data diversity in FL.
• To improve the accuracy of local models in diverse federated environments.

Main Methods:

• FedBranched utilizes probabilistic methods and hidden Markov model (HMM) clustering.
• Client branching is determined by data diversity and performance metrics (MAPE).
• Clustering is based on Euclidean distances of Mean Absolute Percentage Errors (MAPE) from clients.

Main Results:

• The framework was tested on substation energy data for short-term load forecasting using an artificial neural network (ANN).
• FedBranched employed two clustering rounds, creating two distinct client branches with individual global models.
• A significant increase in average MAPE across clients was observed, with one client showing an 11.36% improvement.

Conclusions:

• FedBranched effectively handles client data diversity in federated learning.
• The proposed framework enhances the accuracy of local models, particularly for clients with anomalous behavior.
• This approach shows promise for improving performance in federated learning applications like energy load forecasting.