Development of a new approach to disaggregation and classification
of electrical charges in non-intrusive monitoring systems.
Measurement of Load, Energy Efficiency, NILM, Energy Consumption, Power
Monitoring.
This thesis focuses on energy disaggregation based on a NILM (Non-Intrusive Load Monitoring)
system and aims to investigate and explore methodologies not yet applied, as well as how to
develop an approach to solve the problem of separation of consumption in the expenses of various
equipment, connected to the electrical circuit of the house. The current literature is composed
continued work on the classification of devices by models for cargo breakdown. Such methods
were created from computational models. event identification and pattern recognition. However,
the studies studied still demonstrate great need to reduce computational effort, minimize errors
through optimal pattern recognition processes, override override and distinguishing handsets
with similar signatures. Therefore, one has not yet been found. distinct set of characteristics
able to accurately describe each device. Therefore, This paper seeks to reinforce the demand
for these characteristics. To do this, a rule is introduced of stable state identification, as well as
its mathematical demonstration, used for the step-change recognition on active (P), reactive (Q)
and power factor signals (FP). The identified step-changes define the equipment signatures later
used by three classification methods for recognition of electrical appliances, namely: k-nearest
neighbors algorithm (KNN), Support Vector Machine (SVM) and Genetic Algotithm (GA). Also,
a method based on theWavelet Shrinkage technique for information extraction from the aggregate
signal. The proposed method starts from the idea that several segments can be parsed by distinct
wavelet functions, so the disaggregation of energy can be interpreted as a source separation
problem when only one signal of mixing is known. Therefore, we analyzed the performance of
font modeling with multidimensional vectors and corresponding factorization method. Assuming
that one multidimensional vector composed by the expenditure data of the various equipaments
of the house can be defined, its non-negative factorization is performed in order to extract the
most relevant. The resulting factors are incorporated into the source inference process, in which
only household aggregate consumption is available. The computational experiments showed the
effectiveness of the proposed model to distinguish the different loads under study in relation to
approaches studied so far.