一个面向电力计量系统的联邦学习框架

A Federated Learning Framework for Power Grid Metering System

大数据时代背景下,不同的数据拥有者之间存在信息孤岛的问题,要想得到性能较好的模型,数据必须整合在一起,这经常带来信息安全和数据隐私保护问题。考虑到电力计量系统中各个数据拥有者之间也存在信息安全等问题而无法整合在一起的情况,为了充分利用这些数据完成针对电力计量系统业务应用模型的训练,提出一种基于联邦学习(federated learning,FL)的分布式训练框架,并利用该框架对电力计量数据进行分析。所提出框架旨在保证各个本地电力数据信息安全的前提下,构建去中心化数据的集合以及联合多方数据,进而利用加密后的中间参数,完成多数据源对于联合模型的共同训练。最后通过对联邦学习框架在电力计量领域3项实验结果的分析,证明了该框架的实用性和可行性。

In age of big data,barrier commonly exists between data owners.In order to improve the performance of trained model,it is necessary to aggregate all individual user data,which cause the security and privacy problem.Considering the same situation of power grid metering system,the local data owner always keeps its security and privacy,it is impossible to aggregate them and use all data to train a central model.In this paper,we proposed a distributed framework based on federated learning(FL)concept to train power grid application model by analyzing and using the decentralized power metering data.It’s aimed to train a common model without revealing security and privacy of multi-source local data by using encrypted training parameters.At last,through the analysis of three typical experiment results,this paper proved the usability and feasibility of the FL framework for power grid metering system.