Adversarial Data Anomaly Detection and Calibration for Nonintrusive Load Monitoring

With the increasing prevalence and evolvement of ongoing technologies in household smart meters, nonintrusive load monitoring (NILM) becomes a convenient and cost-effective solution for appliance-level energy monitoring and analysis. As a vital tool for more informed electricity usage management, NILM is naturally intersected with the emerging field of deep learning. However, the predominant focus in deep learning-based NILM research is on enhancing the structure of neural networks, giving little emphasize on the data quality problem which heavily influences the accuracy and robustness of NILM. To address this issue, this article analyses the point-type and pattern-type abnormal data that may affect the accuracy of NILM, and proposes a data anomaly detection method accordingly based on an enhanced cycle-consistent generative adversarial network, which can represent the characteristics of load series in a low-dimensional latent space. Furthermore, by this latent space, we present an anomaly calibration approach based on clustering and nearest neighbors approximation. Extensive experiments using the open data sets are designed to demonstrate the effectiveness of the proposed method in improving NILM accuracy in both on/off states identification and appliance-level load prediction.