Federated two-edge graph attention network with weighted global aggregation for electricity consumption demand...

Area of Science:

• Smart Grid Technology
• Artificial Intelligence
• Data Science

Background:

• Accurate electricity demand forecasting is vital for smart grid stability and preventing failures.
• Federated learning (FL) offers privacy-preserving solutions for data scarcity in electricity consumption data.
• Existing FL methods face challenges with overfitting peak demand variations and over-smoothing predictions due to parameter inconsistencies.

Purpose of the Study:

• To propose a novel federated learning framework, FapDGN, for enhanced electricity demand forecasting.
• To address limitations of existing FL approaches, specifically overfitting and over-smoothing in demand prediction.
• To develop a privacy-preserving method that accurately models both temporal patterns and peak variations in electricity usage.

Main Methods:

• A Federated Two-Edge Graph Attention Network with Weighted Global Aggregation (FapDGN) is proposed.
• The framework constructs hybrid feature representations using temporal and numerical structure edges in graphs.
• A multi-scale attention mechanism captures temporal trends, while dynamic covariance models peak variations to mitigate overfitting.
• A similarity-based adaptive dynamic fusion mechanism is used for server-level parameter aggregation to combat over-smoothing.

Main Results:

• FapDGN effectively models both temporal dynamics and peak variations in electricity consumption.
• The two-edge graph structure and Gaussian distribution modeling mitigate overfitting risks associated with peak demand.
• The adaptive fusion mechanism successfully reduces over-smoothing issues in the global model.
• Experimental results demonstrate FapDGN's superior performance compared to conventional FL methods in electricity demand forecasting.

Conclusions:

• FapDGN offers a robust and privacy-preserving solution for accurate electricity demand forecasting in smart grids.
• The proposed method effectively balances the modeling of temporal patterns and peak variations, overcoming key limitations of existing FL techniques.
• FapDGN demonstrates significant potential for improving the operational efficiency and stability of smart grids.