Inferring Socio-Demographic Information Using Smart Meter Data by Transfer Learning

Abstract

This paper proposes a framework for inferring socio-demographic information using smart meter data. Socio-demographic information can be used to provide effective demand response programs and personalized services. Accordingly, research has been conducted to infer such information using electricity usage patterns which are collected by smart meters. However, collecting household characteristics information and corresponding smart meter data requires considerable effort and cost, making it difficult to obtain sufficient training data. Therefore, in this paper, we present a transfer learning methodology using datasets collected from different countries or regions. In the proposed framework, both the source dataset and target dataset are used to generate a typical daily load profile. The extracted daily load profiles are then used for instance selection step to prevent negative transfer. Also, to improve the performance of the transfer learning model, potentially noisy features are removed. The pre-trained deep learning model is then fine-tuned by the target dataset. Using the proposed method, the information-inferring performance is improved in classification accuracy, F1 score and area under the curve (AUC) metrics.