Resident Change Detection Using Electricity and Gas Metering Data

As smart meter technology advances, advanced metering infrastructure (AMI) enables the collection of high-resolution energy consumption data, such as 15-minute interval electricity and gas usage. While anomaly detection studies have primarily focused on industrial infrastructure, the challenge of identifying specific residential events, such as resident changes, in an unsupervised manner remains largely unaddressed. In this paper, we propose an unsupervised framework to detect the month of a resident change by interpreting significant prediction errors from a deep learning model as anomalies. Our framework utilizes a gated recurrent unit and fully connected layer (GRU-FC) model to predict both monthly consumption patterns and amounts. For training and validation, we used electricity and gas datasets collected from 750 households in Seoul, Republic of Korea, using AMI. These datasets contain ground-truth labels for the actual resident change dates that occurred in the households in 2024. To prepare the datasets for training the GRU-FC model, an efficient outlier removal scheme was developed based on vector quantization. Combined-loss and complementary decision schemes were also employed to jointly leverage pattern and amount features for both electricity and gas data. When only electricity data was used, the detection accuracy was 80.37%, which increased significantly to 89.19% when electricity and gas datasets were used together. This study not only demonstrates the practical feasibility of detecting resident changes but also highlights the effectiveness of a multi-source, prediction-based approach for general time-series anomaly detection.