Lukas Fabri, Daniel Leuthe, Lars-Manuel Schneider, Simon Wenninger
{"title":"Fostering non-intrusive load monitoring for smart energy management in industrial applications: an active machine learning approach","authors":"Lukas Fabri, Daniel Leuthe, Lars-Manuel Schneider, Simon Wenninger","doi":"10.1186/s42162-025-00517-5","DOIUrl":null,"url":null,"abstract":"<div><p>Non-intrusive load monitoring (NILM) is a promising and cost-effective approach incorporating techniques that infer individual applications' energy consumption from aggregated consumption providing insights and transparency on energy consumption data. The largest potential of NILM lies in industrial applications facilitating key benefits like energy monitoring and anomaly detection without excessive submetering. However, besides the lack of feasible industrial time series data, the key challenge of NILM in industrial applications is the scarcity of labeled data, leading to costly and time-consuming workflows. To overcome this issue, we develop an active learning model using real-world data to intelligently select the most informative data for expert labeling. We compare three disaggregation algorithms with a benchmark model by efficiently selecting a subset of training data through three query strategies that identify the data requiring labeling. We show that the active learning model achieves satisfactory accuracy with minimal user input. Our results indicate that our model reduces the user input, i.e., the labeled data, by up to 99% while achieving between 62 and 80% of the prediction accuracy compared to the benchmark with 100% labeled training data. The active learning model is expected to serve as a foundation for expanding NILM adoption in industrial applications by addressing key market barriers, notably reducing implementation costs through minimized worker-intensive data labeling. In this vein, our work lays the foundation for further optimizations regarding the architecture of an active learning model or serves as the first benchmark for active learning in NILM for industrial applications.</p></div>","PeriodicalId":538,"journal":{"name":"Energy Informatics","volume":"8 1","pages":""},"PeriodicalIF":0.0000,"publicationDate":"2025-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://energyinformatics.springeropen.com/counter/pdf/10.1186/s42162-025-00517-5","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Energy Informatics","FirstCategoryId":"1085","ListUrlMain":"https://link.springer.com/article/10.1186/s42162-025-00517-5","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"Energy","Score":null,"Total":0}
引用次数: 0
Abstract
Non-intrusive load monitoring (NILM) is a promising and cost-effective approach incorporating techniques that infer individual applications' energy consumption from aggregated consumption providing insights and transparency on energy consumption data. The largest potential of NILM lies in industrial applications facilitating key benefits like energy monitoring and anomaly detection without excessive submetering. However, besides the lack of feasible industrial time series data, the key challenge of NILM in industrial applications is the scarcity of labeled data, leading to costly and time-consuming workflows. To overcome this issue, we develop an active learning model using real-world data to intelligently select the most informative data for expert labeling. We compare three disaggregation algorithms with a benchmark model by efficiently selecting a subset of training data through three query strategies that identify the data requiring labeling. We show that the active learning model achieves satisfactory accuracy with minimal user input. Our results indicate that our model reduces the user input, i.e., the labeled data, by up to 99% while achieving between 62 and 80% of the prediction accuracy compared to the benchmark with 100% labeled training data. The active learning model is expected to serve as a foundation for expanding NILM adoption in industrial applications by addressing key market barriers, notably reducing implementation costs through minimized worker-intensive data labeling. In this vein, our work lays the foundation for further optimizations regarding the architecture of an active learning model or serves as the first benchmark for active learning in NILM for industrial applications.