中国电机工程学报2026，Vol.46Issue(3)：942-956,中插7,16.DOI:10.13334/j.0258-8013.pcsee.241994
配电网量测数据动态联邦学习框架、自适应隐私保护模型和边缘侧贡献度评估
Dynamic Federated Learning Framework,Adaptive Privacy Protection Model and Edge-side Contribution Assessment for Distribution Network Measurement Data
摘要
Abstract
Measurement data,as a crucial operational element for all stakeholders in distribution network management and a cornerstone asset for enterprises,exhibit diverse characteristics such as varying privacy protection requirements among stakeholders and heterogeneous forms of datasets.These characteristics significantly constrain the empowerment potential of measurement data.Federated learning has garnered widespread attention for its ability to address data silo issues.However,traditional federated learning frameworks are plagued by inadequate privacy protection for participant data,decreased model performance due to data heterogeneity,and a lack of effective incentive mechanisms.To tackle these issues,the adaptive privacy-protected dynamic federated learning framework(AP-DFL)is proposed.First,considering the different emphasis of privacy protection for different load types,the sensitivity of the dataset is defined from the two-dimensional perspectives of anonymity and confidentiality.On this basis,the privacy budget for each round of training on the edge side is dynamically adjusted to achieve adaptive local differential perturbation.Then,combined with the global differential perturbation on the main station side,privacy attacks are effectively avoided.Next,a participant contribution assessment model based on matrix decomposition Shapley values is proposed.This model efficiently calculates contribution values through the reconstructed method of value matrix decomposition under sampling.The aggregation weights are adaptively adjusted based on the contribution values to achieve dynamic federated aggregation,thus enhancing the convergence speed of the model under data heterogeneity.Finally,experimental analysis is conducted on this federated learning framework in typical distribution network scenarios,demonstrating its feasibility.关键词
联邦学习/差分隐私/Shapley值/矩阵分解/量测数据Key words
federated learning/differential privacy/Shapley value/matrix factorization/measurement data分类
信息技术与安全科学引用本文复制引用
王路遥,龚钢军,陆俊,杨佳轩,杨超,刘礼,强仁..配电网量测数据动态联邦学习框架、自适应隐私保护模型和边缘侧贡献度评估[J].中国电机工程学报,2026,46(3):942-956,中插7,16.基金项目
国家重点研发计划项目(2022YFB3105100).National Key R&D Program of China(2022YFB3105100). （2022YFB3105100）
