Household profile identification for behavioral demand response: A semi-supervised learning approach using smart meter data

Abstract Accurate household profiles (e.g., house type, number of occupants) identification is the key to the successful implementation of behavioral demand response. Currently, supervised learning methods are widely adopted to identify household profiles using smart meter data. Such methods could achieve promising performance in the case of sufficient labeled data but show low accuracy if labeled data is insufficient or even unavailable. However, the acquisition of accurately labeled data (usually obtained by survey) is very difficult, costly, and time-consuming in practice due to various reasons such as privacy concerns. To this end, a semi-supervised learning approach is proposed in this paper to address the above issues. Firstly, 78 preliminary features reflecting the household profiles information are extracted from both time and frequency domain. Secondly, feature selection methods are introduced to select more relevant ones as the input of the identification model from the preliminary features. Thirdly, a transductive support vector machine method is adopted to learn the mapping relation between the input features and the output household profile identification results. Case study on an Irish dataset indicates that the proposed approach outperforms supervised learning methods when only limited labeled data is available. Furthermore, the impacts of different feature selection methods (i.e., Filter, Wrapper and Embedding methods) are also investigated, among which the wrapper method performs best, and the identification accuracy improves with the increase of data resolution.