A Graph Label Propagation Semi-Supervised Learning-Based Residential User Profiles Identification Method Using Smart Meter Data

Through extensive interaction between residential users and energy systems, deeper insight into the residential user profiles (e.g. number of residents, have children or not) become an inevitable requirement of energy utilities or retailers to help them provide more efficient and personalized services to targeted residential users. In recent years, supervised learning methods indicate promising performance in identifying residential user profiles from smart meter data when labeled training samples are sufficient, while the accuracies are significantly decrease under the circumstances of insufficient or unavailable of labeled training samples. In practice, the collection of labelled samples and labeling massive unlabeled samples are very difficult, costly and time-consuming. How to reduce the labelled cost while maintaining the identification accuracy is an urgent problem. To address this issue, a graph-based label propagation semi-supervised learning-based residential user profiles identification approach using smart meter data is proposed in this paper. Specifically, a complete and comprehensive feature engineering including feature extraction (78 preliminary features from time-domain and frequency-domain) and feature selection (more relevant features to the targeted labels) is firstly implemented. Furthermore, the final selected features are input the graph-based label propagation semi-supervised learning identification model using limited labelled samples to predict residential user profiles. Case study based on an Irish realistic dataset demonstrate that the proposed method outperforms supervised learning methods especially in the case of limited data.