Inferring the number of residents using smart meter data by transfer learning

Abstract

This paper proposed a framework for inferring the number of residents through smart meter data. Household characteristics can be used as important side information in providing effective demand response programs and personalized services. Accordingly, research has been actively conducted to infer consumer characteristics using electricity usage patterns which are collected by smart meters. However, such information requires considerable effort and cost to investigate, making it difficult to obtain sufficient data for training. Therefore, in this paper, we present a transfer learning methodology using datasets collected from different areas. The proposed framework consists of three parts: (i) From time series smart meter data, daily typical load profile is generated. (ii) To prevent negative transfer, samples that may degenerate performance were removed from the source dataset. (iii) To improve the performance of the transfer learning model, feature selection is conducted from daily load profiles and then fine-tuning the pre-trained deep learning model by the target dataset. By the proposed method, the accuracy of identifying the number of residents is improved when the amount of target dataset is scarce.