Semi-Supervised Federated Analytics for Heterogeneous Household Characteristics Identification

The widespread use of smart meters in households paves the way for retailers to understand household patterns through electricity usage data. This insight helps them offer personalized services and create better demand response strategies. However, smart meter data is highly heterogeneous since it is collected by different retailers using various data sampling methods, over different time periods, and from households with distinct characteristics. Additionally, the labels of household characteristics are obtained by questionnaires, which is labor-intensive and time-consuming, leaving much data unlabeled while privacy concerns prevent data sharing among retailers. To address these challenges, we propose a novel Semi-Supervised Federated Analytics approach for Heterogeneous Smart Meter Data (SF-Heter). This method keeps raw data local and exchanges analytics outputs, called prototypes, between retailers and a central server, thus dealing with heterogeneous data and protecting privacy. SF-Heter utilizes a new model structure named MODlinear, which enhances feature extraction through contrastive learning and multi-kernel time-series analysis. Meanwhile, SF-Heter efficiently utilizes unlabeled data by generating high-quality pseudo-labels and prototypes using MODlinear and integrated with a quality-controlled semi-supervised loss mechanism. Extensive tests on the Irish dataset show that SF-Heter effectively handles data heterogeneity and optimizes the use of unlabeled data.