Unknown appliance detection for non-intrusive load monitoring using normalized k-nearest neighbors

Non-intrusive load monitoring (NILM) aims to decompose the aggregated power consumption profiles into those of individual appliances, offering significant potential to enhance energy efficiency. In recent years, machine learning-based load recognition methods have successfully addressed closed-set recognition problems, where training and testing data are drawn from the same known classes. Real-world deployments, however, require open-set recognition: models must cope with unknown or unseen appliances at test time, rendering traditional closed-set approaches less effective. To address this challenge, we propose a simple yet effective open-set load recognition approach based on normalized k-nearest neighbor distances and percentile-based thresholding. Unlike existing open-set NILM methods that rely on deep neural networks, the proposed approach avoids deep learning entirely and employs a simple, data-driven thresholding method. The approach needs only a small set of current-derived features, avoids the complexity of deep learning models, and integrates seamlessly with existing closed-set classifiers. Extensive experiments on two public datasets show that the proposed method achieves performance comparable to or better than state-of-the-art deep neural network methods in both accuracy and efficiency, and achieves F1-scores above 91% for unknown appliance detection and macro-averaged F1-scores in the range of 90%–95% across different open-set scenarios, while maintaining minimal computational and memory overhead. The method has also been deployed on a low-cost embedded platform, demonstrating its practical applicability. Code is available at https://github.com/zhz-yan/L2NC .