Smart plug for real-time energy monitoring and optimization through micro-moments with semi-supervised federated learning

Smart energy management systems face the critical challenge of developing privacy-preserving and personalized energy optimization approaches. In real-world deployments, scalability and practicality are restricted by the reliance on manual labeling for anomaly detection, while centralized approaches frequently compromise user privacy. To mitigate privacy concerns and facilitate adaptive and scalable energy optimization, this study presents a novel smart plug system enabled by a Semi-Supervised Federated Learning (SSFL) framework. The framework is designed to operate in two stages: an initial supervised training phase that utilizes 20 % of the publicly available Dutch Residential Energy Dataset (DRED) dataset to develop a robust global model and a subsequent SSFL training phase that utilizes pseudo-labeling for the remaining of DRED and our Qatar University Dataset (QUD) datasets, which are split among three clients. Decentralized model updates are ensured by FL, which also ensures the security and localization of sensitive user data. For the preliminary evaluation, the system was deployed in three residences for two weeks, collecting high-frequency energy and environmental data. The global model was refined daily via federated updates, leading to consistent improvements in the accuracy of anomalous energy consumption detection. The results highlighted the system’s ability to accommodate a wide range of user energy consumption patterns while safeguarding privacy. The framework also improves scalability and user convenience by eliminating the need for manual intervention, automating labeling via pseudo-labeling.