Advanced classification and identification of plugged-in electric loads
MetadataShow full item record
The total electricity consumption of plugged-in electric loads (PELs) currently accounts for more usage than any other single end-use service in residential and commercial buildings. Compared with other categories of electric loads, PELs possess significant potential to be efficiently controlled and managed in buildings. Therefore, accurate and reliable PEL identification methods that are used to collect identity and performance information are desired for many purposes. However, few existing electric load identification methods are designed for PELs to handle unique challenges such as the diversity within each type of PEL and similarity between different types of PELs equipped by similar front-end power supply units. The objective of this dissertation is to develop non-intrusive, accurate, robust, and applicable PEL identification algorithms utilizing voltage and current measurements. Based on the literature review of almost all existing features that describe electric loads and five types of existing methods for electric load identification, a two-level framework for PELs classification and identification is proposed. First, the supervised self-organizing map (SSOM) is adopted to classify a large number of PELs of different models and brands into several groups by their inherent similarities. Therefore, PELs with similar front-end power supply units or characteristics fall into the same group. The partitioned groups are verified by their power supply unit topology. That is, different groups should have different topologies. This dissertation proposes a novel combination of the SSOM framework and the Bayesian framework. Such a hybrid identifier can provide the probability of an unknown PEL belonging to a specific type of load. Within each classified group by the SSOM, both static and dynamic methods are proposed to distinguish PELs with similar characteristics. Static methods extract steady-state features from the voltage and current waveforms to train different computational intelligence algorithms such as the SSOM itself and the support vector machine (SVM). An unknown PEL is then presented to the trained algorithm for identification. In contrast to static methods, dynamic methods take into consideration the dynamics of long-term (minutes instead of milliseconds) waveforms of PELs and extract elements such as spikes, oscillations, steady-state operations, as well as similarly repeated patterns.