A differential privacy-based personalized federated learning research on photovoltaic power prediction in cloud-edge computing environments

Federated learning has emerged as a promising solution to overcome data isolation in photovoltaic power stations, offering a distributed approach compatible with cloud computing technologies. However, despite Federated Learning can protect privacy, it remains vulnerable to inference attacks during training a global model in cloud servers, potentially leaking some information about participating members and posing serious security risks. Moreover, Federated Learning encounters practical challenges in cloud environment, including data heterogeneity across photovoltaic power stations. To address these issues, we propose a novel photovoltaic power prediction framework based on personalized federated learning and long short-term memory. This decentralized approach enhances prediction accuracy using local station data. We introduce gradient clipping to prevent gradient instability and employ Gaussian-based local differential privacy to counter semi-honest or malicious client attacks. Experimental results show that the long short-term memory network based on the personalized federated learning framework can improve the prediction accuracy compared with the traditional long short-term memory model, while effectively protecting the local data privacy of users and avoiding the risk of data leakage during the transmission process.