Enhancing Consumer Privacy in Federated Load Forecasting Through Single-Layer Aggregation

Consumer energy data present unique challenges in terms of privacy and security due to the vast amounts generated by Internet of Things-based devices. Federated learning (FL) has emerged as a privacy-preserving and cost-efficient approach for training machine learning models, offering solutions to protect consumers’ data from breaches. However, FL models are still vulnerable to privacy attacks. To mitigate these risks, differential privacy is applied to local model weights, enhancing privacy but often reducing prediction accuracy. To address this tradeoff, we propose a novel single-layer federated weight aggregation framework for consumer energy load forecasting. This privacy-aware approach aggregates only the single layer of local models, ignoring other layers to minimize privacy risks. Extensive simulations with real-world consumer energy data show the framework's effectiveness, achieving a mean absolute error (MAE) of 1.13 kWh, close to the baseline of 1.01 kWh. It significantly outperforms traditional FL-based methods (FedAvg), which have an MAE of 4.2 kWh at a noise level (ϵ) of 0.2, while providing 211.09 times better communication cost efficiency.