PPenergyNET: Privacy-Preserving Multi-Energy Load Forecasting in Energy Internet Considering Energy Coupling
Yigong Zhang, Qiushi Cui, Lixian Shi, Jianyu Pan, Jian Li
Abstract
Accurate multi-energy load forecasting (MELF) is critical to the management and operation of energy internet (EI). Current MELF methods gather the data from various energy loads for model training in a centralized manner. However, different energy utilities in EI intend to keep the raw data locally due to privacy reasons, which leads to the data silos of EI. To this end, we propose a PPenergyNET to perform privacy-preserving MELF in a distributed manner and break data silos in EI. Specifically, a ring-structure vertical federated learning is devised to protect vertical partition data privacy, and fix the mismatch that leads to incalculable training loss in the cloud to reconstruct the loss back-propagation for model updating. Then, a split feature extraction method is designed to prevent the characteristics of specific load from being submerged by the data of multi-energy loads to improve forecast resolution. Thirdly, a modified homoscedastic uncertainty based multi-task learning method is proposed to consider the multi-energy coupling with a convergence proof. Numerical results show that PPenergyNET achieves superior trade-offs between privacy protection and forecasting accuracy. More importantly, PPenergyNET contributes a new idea to improve interoperability among different energy systems.