Federated Learning for Energy Load Forecasting in Smart Homes

The increasing complexity of residential energy consumption, driven by climate variability, smart devices, and distributed renewables, demands predictive solutions that are accurate, privacy-friendly, and adaptive. Therefore, forecasting energy demand is essential for suppliers to balance supply and demand in real time. In Latin America, these challenges are intensified by strict data protection laws, heterogeneous infrastructures, and dynamic consumption behaviors. In this context, traditional centralized load forecasting models have proven insufficient, as they lack the flexibility, transparency, and adaptability required for smart home energy management systems operating in distributed, non-stationary, and outage-prone conditions, limiting their usefulness for real-world decision-making. To address these limitations, we propose a generalized, adaptive, and explainable federated learning framework for residential load forecasting. The approach integrates data curation, preprocessing, federated clustering, and a forecasting model based on a temporal convolutional network with global attention, optimized through Bayesian optimization and augmented with feature attribution techniques, ensuring accurate and interpretable predictions. Preliminary results, validated with real-world datasets, show that the proposed system consistently improves accuracy in different scenarios, while providing interpretable information to support energy operators' decision-making. This work contributes to developing more resilient and sustainable energy infrastructures, aligned with the Sustainable Development Goals and the energy sector's digital transformation.