Open Access

Daniel Weißhaar

• This Ph.D. thesis deals with the question of how a device, such as a dishwasher, can be detected in an unknown electronic network based on current and voltage. And does this only apply to dishwashers, or can other appliances also be considered? This involves the development of an algorithm for detecting complex appliances, the development of a hardware and software prototype, and the development of a method for improving future algorithm developments in the research area of Non-Intrusive Load Monitoring. The underlying research area Non-Intrusive Load Monitoring deals, among other things, with the question of how the overall performance can be assigned to individual consumers from a data collection that is as less intrusive as possible. Many other questions may be derived from this, including the politically motivated question of the energy-saving potential. In the work described here, a special aspect is in the foreground. Many algorithms and environments in which NILM is tested require prior knowledge of the appliances used or even the number of appliances in the network. This work is about what an algorithm can look like that dispenses with this a priori knowledge. Based on the current and voltage input signals, information is extracted step by step. Devices in the signal are systematically detected using various algorithms such as event detection, clustering, and a transformation of the event cluster list. Ultimately, complex appliances are identified as a whole, which makes it possible to describe these appliances in their performance profile and to record additional information such as the runtime. The algorithm developed here is compared with another algorithm from the literature. The implementation of such an algorithm on hardware would make it possible in the future to generate a self-learning Non-Intrusive Load Monitoring device. As part of the tri-national Interreg research project Smart Meter Inclusive (SMI), a Smart Meter Inclusive prototype was developed in cooperation between UHA and HFU, which, in addition to normal energy measurement, can also assign the power to the consumers located within the measured electronic network. As part of this promotion, the software for data logging and classification was developed. This software is distributed over several hardware components and their interaction, as well as the implementation of various previously developed algorithms from preliminary work in the form of real-time processing. The prototype includes a machine learning model which can distinguish between different appliances at runtime and identify them reliably. In the area of Non-Intrusive Load Monitoring, there is the problem that data sets that meet certain criteria are required for the development of new algorithms. Generating real data is a laborious and time-consuming activity, and measuring appliances that are rarely switched on and off means handling very large amounts of data. For this purpose, a method was developed to synthesize new data sets from existing real data. The basis is formed by individual measurements in which an appliance was switched on and off many times in a row. Through the artificial generation of switching cycles, these real switching cycles can be extended and new data sets from these individual measurements can be combined using defined criteria. Among other things, it is possible to determine how large the pauses and lengths of the switching cycles should be, as well as the number of devices switched on at a time. For the future development of various NILM algorithms, this method thus offers a possible database for rapid pre-development before these algorithms can later be tested on real data.