Nowadays, smart meters are deployed in millions of residential households to gain significant insights from fine-grained electricity consumption data. The information extracted from smart meter data enables utilities to identify the socio-demographic characteristics of electricity consumers and then offer them diversified services. Traditionally, this task is implemented in a centralized manner with the assumption that utilities have access to all the smart meter data. However, smart meter data are measured and owned by different retailers in the retail market who may not be willing to share their data. To this end, a distributed electricity consumer characteristics identification method is proposed based on federated learning, which can preserve the privacy of retailers. Specifically, privacy-perseverance principal component analysis (PCA) is exploited to extract features from smart meter data. On this basis, an artificial neural network is trained in a federated manner with three weighted averaging strategies to bridge between smart meter data and the socio-demographic characteristics of consumers. Case studies on the Irish Commission for Energy Regulation (CER) dataset verify that the proposed federated method has comparable performance with the centralized model on both balanced and unbalanced datasets.

中文翻译：

---

电力消费者特征识别：联合学习方法

如今，数以百万计的居民家庭部署了智能电表，以从细粒度的用电量数据中获得重要见解。从智能电表数据中提取的信息使公用事业公司能够识别电力消费者的社会人口特征，然后为他们提供多样化的服务。传统上，该任务以集中方式实施，假设公用事业公司可以访问所有智能电表数据。然而，智能电表数据由零售市场中的不同零售商测量和拥有，他们可能不愿意共享他们的数据。为此，提出了一种基于联邦学习的分布式电力消费者特征识别方法，可以保护零售商的隐私。具体来说，隐私持久性主成分分析（PCA）被用来从智能电表数据中提取特征。在此基础上，人工神经网络通过三种加权平均策略以联合方式进行训练，以在智能电表数据与消费者的社会人口特征之间架起桥梁。爱尔兰能源监管委员会 (CER) 数据集的案例研究证实，所提议的联合方法在平衡和非平衡数据集上的性能与集中式模型相当。