Privacy-Preserving Federated Learning for Value-Added Service Model in Advanced Metering Infrastructure

• Published in: IEEE transactions on computational social systems 2024-02, Vol.11 (1), p.1-15
• Main Authors: Zhang, Xiao-Yu, Cordoba-Pachon, Jose-Rodrigo, Guo, Peiqian, Watkins, Chris, Kuenzel, Stefanie
• Format: Article
• Language: English
• Subjects: Accuracy; Advanced metering infrastructure; Advanced metering infrastructure (AMI); attention-based deep learning; Cloud computing; Computational modeling; Consumers; Customer satisfaction; Data models; differential privacy (DP); energy cyber-physical social system (CPSS); Federated learning; federated service; Internet of Things; Load modeling; Neural networks; Privacy; Random noise; Servers; Smart grid; Smart grids; User satisfaction
• ISSN: 2329-924X; 2329-924X; 2373-7476
• DOI: 10.1109/TCSS.2022.3204361

Advanced metering infrastructure (AMI) is the backbone of the next generation smart city and smart grid; it not only provides near real-time two-way communication between the consumers and the energy systems, but also enables the third parties (TPs) to provide relevant value-added services to the consumers to improve user satisfaction. However, the existing services are implemented in a centralized manner, which has potential and associated security and privacy risks also increased with Internet-of-Things (IoT) devices. To better balance the quality of the services and ensure users' privacy, a TP AMI service model based on differentially private federated learning (FL) is proposed in this article. Instead of sending the private energy data to the cloud server, the proposed service model trains the neural network models locally, and only model parameters are shared with the central server. Moreover, the identity of individuals is eliminated by adding random Gaussian noise during the secure aggregation. Furthermore, an attention-based bidirectional long short-term memory (ATT-BLSTM) neural network model is adopted to solve the long-range dependency problem of conventional neural networks. In the case study, a residential short-term load forecasting (STLF) task is implemented to evaluate the performance of the proposed model. Compared with other state-of-the-art energy service models, the proposed one can achieve similar accuracy as the typical centralized model and balances the trade-off between privacy loss and prediction accuracy flexibly.