Abstract
The shift towards solar power underscores the importance of managing the variability of renewable energy. While smart energy systems play a crucial role by tracking real-time energy usage and adjusting to the unpredictability of renewables to maintain balance, their integration with cloud-based predictive models for improved energy management raises privacy and security concerns [1]. Federated learning [2] addresses these by facilitating local training of models at the client side (edge device) and sharing updates globally by aggregating the local models into a global one (cloud server), offering a method to preserve privacy, though challenges persist with cloud-based inference. Meanwhile, edge computing [3] and encrypted inference [4] emerge as alternatives, each facing limitations in computation and data protection during training. In this research, we develop a deep learning-based framework for privacy-preserving prediction of household energy consumption and photovoltaic (PV) generation, ensuring data privacy during the training and inference phases. Utilizing federated learning, our framework constructs local models on the prosumer side, which predict both electricity household consumption and PV power generation. These local models are then combined into global models accessible to the entire community. To ensure privacy during inference, we utilize Homomorphic Encryption (HE), specifically the Brakerski/Fan-Vercauteren (BFV) [5] scheme, allowing for the processing of encrypted data on cloud services. User data is encrypted on the client's mini-computer (edge device) and then sent to the cloud along with the public encryption key and encoding parameters, for inference. Subsequently, the cloud service sends the encrypted results back to the user, where they are decrypted on the client's edge device using the private encryption key. Such a technique will preserve the privacy of the data because encrypted data will never leak any information. A key aspect of our approach is balancing the performance of local models with the computational demands of the global model operating on encrypted data. This balance is critical to achieving good predictive performance while staying within the constraints of the Noise Budget (NB) allowed by the BFV encryption scheme. We follow a strict design process, iteratively refining the local and global models to ensure they deliver optimal performance without exceeding the NB. This iterative process continues until the models meet our stringent criteria for efficiency and privacy preservation, demonstrating our commitment to advancing predictive models technologies while protecting the user data. This solution notably addresses one of the biggest issues in classic predictive models – data privacy – while still maintaining strong performance of deep learning model. By balancing these aspects, our framework presents a comprehensive, privacy-focused approach for household energy management. On the other hand, the direct implementation of this technique within smart grids is slowed down by persisting technical challenges. The integration of Homomorphic Encryption (HE) and shared predictive models (operations on encrypted data from different users), which remains unresolved, poses significant obstacles to their immediate application in smart energy systems. The key issue arises from users having unique encryption settings. A possible solution is Threshold Cryptography, where a shared public key is used with divided private keys among users. This requires a set number of users to jointly decrypt data. Despite its complexity and being under research, it could be explored further in future project developments. The preliminary results showed that our proposed technique, which fully preserves privacy, still performs well compared to predictive models that compromise data privacy, with less than a 2% decrease in accuracy. Additionally, introducing smart energy management systems that offer both consumption and PV power predictions proves very beneficial for prosumers, aiding them in more effectively managing their energy systems.