A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Non-Intrusive Load Monitoring (NILM): combining multiple distinct electrical features and unsupervised machine learning techniques
Common electricity meters measure only the overall energy consumption. To assist in energy savings and to enrich smart home applications with energy data, a detailed breakdown is necessary. Non-Intrusive Load Monitoring (NILM) analyzes the overall electrical signal and separates it into its components, by identifying specific electrical signatures. So far most studies focused on low frequency NILM since standard measurement hardware can be used. In addition some specialized research has been carried out in the high frequency field. The goal of this work is to examine various electrical features and combine them in a holistic manner. Therefore we analyzed low, mid and high frequency features like active and reactive power, harmonics and line-conducted electromagnetic interference signals for their applicability in NILM. In a next step we developed an unsupervised learning algorithm which uses the most promising features to disaggregate the overall load profile. One property of our machine learning algorithm is the detection of also completely unknown devices. Another is the ability of real-time analysis of the meter data. Our results showed that higher frequency features can assist significantly in the task of load disaggregation. Harmonics for example can be especially beneficial by separating devices with a similar active power intake. Line-conducted disturbances - a subclass of electromagnetic interference signals - on the other hand can be useful to trace variable loads or to split devices of the same model. Still some more work has to be carried out in the field of overlapping distortions. Our algorithm evaluation showed some promising results for a privately recorded and publically available dataset. In addition industrial measurements were examined resulting in a high event detection performance. With this work we show the potential of combining low, mid and high frequency features. For future NILM algorithms the benefits of a high sampling rate should be considered and ideally be integrated into the smart ...
Non-Intrusive Load Monitoring (NILM): combining multiple distinct electrical features and unsupervised machine learning techniques
Common electricity meters measure only the overall energy consumption. To assist in energy savings and to enrich smart home applications with energy data, a detailed breakdown is necessary. Non-Intrusive Load Monitoring (NILM) analyzes the overall electrical signal and separates it into its components, by identifying specific electrical signatures. So far most studies focused on low frequency NILM since standard measurement hardware can be used. In addition some specialized research has been carried out in the high frequency field. The goal of this work is to examine various electrical features and combine them in a holistic manner. Therefore we analyzed low, mid and high frequency features like active and reactive power, harmonics and line-conducted electromagnetic interference signals for their applicability in NILM. In a next step we developed an unsupervised learning algorithm which uses the most promising features to disaggregate the overall load profile. One property of our machine learning algorithm is the detection of also completely unknown devices. Another is the ability of real-time analysis of the meter data. Our results showed that higher frequency features can assist significantly in the task of load disaggregation. Harmonics for example can be especially beneficial by separating devices with a similar active power intake. Line-conducted disturbances - a subclass of electromagnetic interference signals - on the other hand can be useful to trace variable loads or to split devices of the same model. Still some more work has to be carried out in the field of overlapping distortions. Our algorithm evaluation showed some promising results for a privately recorded and publically available dataset. In addition industrial measurements were examined resulting in a high event detection performance. With this work we show the potential of combining low, mid and high frequency features. For future NILM algorithms the benefits of a high sampling rate should be considered and ideally be integrated into the smart ...
Non-Intrusive Load Monitoring (NILM): combining multiple distinct electrical features and unsupervised machine learning techniques
Bernard, Timo (author) / Fuhr, Norbert
2018-07-17
Theses
Electronic Resource
English
big data , artificial intelligence , high frequency electrical features , energy efficiency , Fakultät für Ingenieurwissenschaften » Informatik und Angewandte Kognitionswissenschaft , device identification , energy consumption , energy management , load disaggregation , Non-Intrusive Load Monitoring (NILM) , ddc:620 , smart metering , load management , unsupervised machine learning