TY - GEN

AU - Völker, Benjamin

AU - Albert-Ludwigs-Universität Freiburg im Breisgau Lehrstuhl für Rechnerarchitektur

AU - Albert-Ludwigs-Universität Freiburg Fakultät für Angewandte Wissenschaften

AU - Scholl, Philipp Marcel

AU - Becker, Bernd

AU - Scholl, Christoph

AU - Laerhoven, Kristof Van

TI - Acquiring, congregating, and processing high-frequency electricity data in distributed environments

PB - Universität

KW - Electricity

KW - Non-Intrusive Load Monitoring

KW - Electricity Monitoring

KW - Appliance Classification

KW - Data Acquisition Systems

KW - (local)doctoralThesis

PY - 2022

N2 - Abstract: The energy consumption of a home depends on the behavior of its inhabitants offering a promising energy saving potential. However, this potential can only be unfolded to full extent if the consumption of each individual appliance is know. Non-Intrusive Load Monitoring (NILM) offers a retrospective way to get individual appliance consumption data. If such data are combined with eco-feedback techniques it can help to better understand a user’s electricity usage to ultimately save energy. Researching NILM algorithms, and in particular, the development of the underlying supervised machine learning techniques, requires adequate datasets with corresponding ground truth data and methodologies to create more labeled data when needed. Adding detailed labels to such datasets is a time-consuming and error-prone process. Deploying a supervised NILM system typically requires a dedicated system training procedure hampering their widespread adoption. The thesis at hand presents several strategies to address these challenges in order to improve the adoption of NILM. In particular, a set of requirements is presented for acquiring and congregating high-frequency electrical measurements in distributed environments. These are handled by a novel recording framework comprised of a central recording director and two prototype Data Acquisition Systems (DAQs), one for aggregated and one for plug-level data. The developed methodologies allow the DAQs to deliver highly accurate and time-synchronized data while using rather inexpensive components. To add precise and descriptive labels to such data, a semi-automatic labeling method is developed and evaluated on two publicly available datasets. The method improves the labeling efficiency up to 74% and has been integrated into a novel labeling tool implemented as a web-application. The framework and labeling tool have been used to collect and label the Fully-labeled hIgh-fRequency Electricity Disaggregation (FIRED) dataset. It contains 101 days of 8 kHz aggregated current and voltage measurements of the 3-phase electricity supply of a typical residential apartment in Germany. The data also includes synchronized 2 kHz plug-level readings of 21 individual appliances, other environmental sensor measurements, and descriptive event labels of all appliances, resulting in a complete and versatile residential electricity dataset. Furthermore, several domain specific features and classifiers are evaluated regarding their suitability for (event-based) NILM targeted for resource constrained systems. Data cleaning methods are evaluated which remove the steady-state energy consumption of other appliances from the aggregated data of a given appliance event. As plug-level data delivers less noisy individual appliance data, it is shown that the inclusion of such data during training results in a performance gain for appliance classification algorithms. Finally, a novel supervised NILM system is proposed and evaluated which uses a combination of aggregated and individual appliance data to improve and aid the training process while only requiring minimal user interaction

CY - Freiburg

UR - http://slubdd.de/katalog?TN_libero_mab2

ER -