DPWGAN: High-Quality Load Profiles Synthesis With Differential Privacy Guarantees
Jiaqi Huang, Qiushi Huang, Gaoyang Mou, Chenye Wu
Abstract
Smart meters have collected massive amounts of fine-grained load data from users, enabling various load profile analyses that can help improve the efficiency of smart grids. However, the smart meter data may leak private information, raising public concerns. To address this issue, current approaches typically employ data perturbation mechanisms or data generation mechanisms to ensure privacy when analyzing load profiles, but these approaches are either inflexible or not guaranteed to mitigate the leakage issue. To this end, we propose a differentially private Wasserstein Generative Adversarial Networks (DPWGAN) approach in this study. This approach can privately convert a real-world dataset into a high-quality synthetic load dataset so that studies and analyses conducted on the synthetic dataset can automatically satisfy user-level differential privacy guarantees. The extensive numerical studies highlight that our approach acts as an excellent substitute for the original dataset in real-world load profiling tasks.