Spatiotemporal Federated Learning for Privacy-Preserving Load Forecasting and Appliance Scheduling in Smart City Homes

As smart cities evolve, optimizing energy use within residential environments via intelligent consumer electronics becomes crucial to ensure grid stability, enhance user comfort, and promote sustainability. This paper proposes a novel IoT-enabled demand response (DR) framework that integrates AI-based spatio-temporal load forecasting and intelligent appliance scheduling through federated learning (FL) across consumer electronics such as smart meters, home hubs, and smartphones. The system leverages household-level energy consumption data, which are inherently temporal and spatial in nature, to make localized, context-aware scheduling decisions while preserving data privacy. To address key challenges in home energy management, such as data heterogeneity, forecast uncertainty, and multi-objective optimization, the framework employs a Split Federated Learning (Split-FL) architecture with Gated Recurrent Units (GRU) for accurate, privacy-preserving load forecasting. Raw data remain on edge devices, ensuring compliance with privacy regulations while enabling collaborative model training. Additionally, a robust hybrid metaheuristic algorithm, combining the Improved Grey Wolf Optimizer (IGWO) and Teaching-Learning-Based Optimization (TLBO), is deployed at the Fog layer to solve the complex multi-objective appliance scheduling problem, explicitly accounting for forecast uncertainties modeled via ellipsoidal sets. Comprehensive experimental evaluations in real-world datasets (REDD, UK-DALE and Pecan Street) demonstrate that the proposed framework achieves an average reduction of 32% in total energy cost, a decrease of 51% in user discomfort, and a reduction of 40% in the peak to average ratio, while maintaining a compliance rate of demand response 92%. The Split-FL GRU forecasting model also outperforms centralized approaches in prediction accuracy. By enabling decentralized, adaptive, and privacy-preserving intelligence across heterogeneous consumer devices, this framework supports scalable and resilient energy optimization in smart homes. It directly contributes to the operational efficiency and sustainability of smart city infrastructures and lays the foundation for future location-sensitive energy recommender systems.