Adaptive Single-layer Aggregation Framework for Energy-efficient and Privacy-preserving Load Forecasting in Heterogeneous Federated Smart Grids

Abstract

Federated Learning (FL) enhances predictive accuracy in load forecasting by integrating data from distributed load networks while ensuring data privacy. However, the heterogeneous nature of smart grid load forecasting introduces significant challenges that current methods struggle to address, particularly for resource-constrained devices due to high computational and communication demands. To overcome these challenges, we propose a novel Adaptive Single Layer Aggregation (ASLA) framework tailored for resource-constrained smart grid networks. The ASLA framework mitigates data heterogeneity issues by focusing on local learning and incorporating partial updates from local devices for model aggregation in adaptive manner. It is optimized for resource-constrained environments through the implementation of a stopping criterion during model training and weight quantization. Our evaluation on two distinct datasets demonstrates that quantization results in a minimal loss function degradation of 0.01% for Data 1 and 1.25% for Data 2. Furthermore, local model layer optimization for aggregation achieves substantial communication cost reductions of 829.2-fold for Data 1 and 5522-fold for Data 2. The use of an 8-bit fixed-point representation for neural network weights leads to a 75% reduction in storage/memory requirements and decreases computational costs by replacing complex floating-point units with simpler fixed-point units. By addressing data heterogeneity and reducing storage, computation, and communication overheads, the ASLA framework is well-suited for deployment in resource-constrained smart grid networks.