A Federated Learning Approach with Spiking Neurons and Clustering for Smart Meter Data

Residential smart meters data is essential for load flow analysis and demand-supply balance in smart grids. However, centralized load forecasting methods compromise consumer privacy and lead to high communication overhead due to the transmission of sensitive data. Federated learning (FL) addresses these concerns by enabling decentralized model training. However, traditional FL models rely on computationally intensive architectures, such as Long Short-Term Memory (LSTM) networks, which are not suitable for resource-constrained devices. Additionally, a single global FL model fails to account for diverse household consumption patterns, resulting in suboptimal forecasting performance. To overcome these challenges, we propose a novel federated averaging (FedAvg) approach based on integrate-and-fire (IF) spiking neurons for short-term load forecasting (STLF). Spiking Neural Networks (SNNs) provide event-driven, energy-efficient processing, reducing computational complexity and communication costs. In addition, we introduce a parameter-based clustering mechanism that leverages the statistical properties of model updates to group households with similar consumption patterns. Our method enhances model aggregation and effectively captures user-specific consumption patterns. Experimental results on a real-world dataset of 18 homes demonstrate that our FedAvg-IF model reduces RMSE by approximately 74% and MAE by 82% compared to the FedAvgLSTM model. The introduction of the clustering mechanism (FedAvg-IF-C) further decreases communication costs by about 89% and computational time by 61%, respectively.