Heuristic based federated learning with adaptive hyperparameter tuning for households energy prediction

Area of Science:

• Artificial Intelligence
• Energy Systems
• Machine Learning

Background:

• Federated Learning (FL) enables AI model training on edge devices for electrical load forecasting.
• Non-IID data in FL degrades prediction accuracy, necessitating adaptive hyperparameter optimization.
• Existing FL methods struggle with heterogeneous household energy consumption patterns.

Purpose of the Study:

• To develop a novel hierarchical federated learning solution for efficient electrical load forecasting.
• To improve prediction accuracy by addressing non-IID data challenges through adaptive hyperparameter tuning.
• To optimize model aggregation and reduce computational load on edge devices.

Main Methods:

• Clustering households by energy profiles at the edge and aggregating at the fog level.
• Hierarchical simulated annealing for optimizing federated model aggregation, prioritizing high-performing models.
• Genetic algorithm-based hyperparameter optimization to reduce edge node computational burden.
• Evaluating prediction accuracy against federated averaging.

Main Results:

• Significant improvement in average prediction accuracy for household energy patterns.
• Demonstrated ability to capture complex energy consumption dynamics effectively.
• Network traffic impact maintained below 30 KB across different network layers.
• Reduced model update size and communication rounds by 30% through hyperparameter tuning.

Conclusions:

• The proposed hierarchical federated learning approach enhances electrical load forecasting accuracy.
• Adaptive hyperparameter tuning is crucial for improving FL performance with non-IID energy data.
• The solution offers efficiency gains, particularly beneficial for resource-constrained edge environments.