Pattern Consistency Learning-Enabled Nonintrusive Load Monitoring Considering Limited Appliance Annotations

Deep learning (DL)-based nonintrusive load monitoring (NILM) methods face significant challenges due to the high cost of acquiring appliance annotations as training labels. While existing semisupervised learning (SSL) methods leverage low-cost unlabeled data to enhance performance, they typically depend on high-quality unlabeled data and involve complex operation processes. To address them, this article proposes a pattern consistency learning (PCL)-enabled SSL method that reduces reliance on strict unlabeled data requirements while delivering superior monitoring performance. Specifically, a tailored collaborative training manner is developed to train a single network using both labeled and unlabeled data, leveraging structure characteristics from unlabeled data and simultaneously capturing task-specific patterns from labeled data. This enables the proposed method to learn essential general patterns of aggregated load. Meanwhile, a multihead partitioned convolutional network (MPCN) is designed to extract the key features of aggregated load through multiview temporality mixture, facilitating high-precision predictions within the designed SSL framework. Comprehensive results demonstrate the superiority of the proposed method in NILM, particularly in scenarios with limited appliance annotations, as well as its efficiency with a reduced computational burden compared to existing SSL-based methods.