A Federated Deep Reinforcement Learning Framework for Real-Time Demand Response Optimization in IIoT-based Smart Home Automation

Abstract

The residential market has seen the rapid take-up of Internet of Things (IoT) devices, connected IIoT smart systems due to it. This situation gives rise to new issues and challenges for the energy management implemented in real time. The traditional centralized energy management mechanisms face great challenges like excessive latency, confidentiality vulnerability, scalability issues across heterogeneous smart home networks. The paper proposes a Federated Deep Reinforcement Learning (D-RLR) framework based on DQN agents and federated learning method. The framework allows for the decentralized training of DQN models in smart home nodes with the restriction that raw energy consumption data remain local. Each agent carries out the processing of grid price signal, usage data of appliances, observed amount and availability of local renewable energy to develop optimal load scheduling policy. The overall demand response policy is obtained by periodic federated aggregation of local models, helping to reduce communication latency. Experiments conducted on a simulated IIoT smart home environment, constructed using appliance consumption profiles derived from publicly available residential energy datasets and realistic solar irradiance-based renewable generation profiles, demonstrate that the proposed framework reduces peak energy consumption by 23% and improves demand response efficiency by 18% relative to centralized DRL baselines. Communication overhead generates roughly 65% lesser and user privacy is preserved throughout by design. The mechanism scales across nodes and supports real-time scheduling decisions under dynamically changing grid conditions. Therefore, it can be deployed practically for next-generation residential energy management.