Nombre: ALOISIO RAMOS DA PAIXÃO
Tipo: MSc dissertation
Fecha de publicación: 01/12/2016
Supervisor:
Nombre |
Rol |
|---|---|
| WANDERLEY CARDOSO CELESTE | Advisor * |
Junta de examinadores:
| Nombre |
Rol |
|---|---|
| ANIBAL COTRINA ATENCIO | Internal Examiner * |
| HELDER ROBERTO DE OLIVEIRA ROCHA | External Examiner * |
| LUIS OTAVIO RIGO JUNIOR | Co advisor * |
| WANDERLEY CARDOSO CELESTE | Advisor * |
Resumen: This research aims to development a system of intelligent classification of similar and
non similar electrical loads, using non-intrusive measurement for the acquisition of
voltage and current electric signals. Initially an experimental platform containing an
arrangement with 4 similar electrical loads, that is, of the same manufacturer and with
identical technical specifications, is implemented. Subsequently, an arrangement with
4 non-similar electric loads is used, in order to allow a comparison with the works
observed in the recent literature. Six intelligent classifiers are used in the identification
process, namely: k-means, Case-Based Reasoning (CBR), CBR+k-means and three
Multi-Layer Perceptron (MLP) Artificial Neural Networks (ANN), being one ANN with 4
neurons in the hidden layer (MLP-4) and two ANNs with 8 neurons in the hidden layer
(MLP-8 and MLP-8-C30000) that differ only in the number of cycles used as criterion
of Learning. The experiments are performed using electrical signals sampled in the
frequencies of 6.25kHz, 12.5kHz and 25kHz, in order to verify the influence of the
sampling rate on the identification process. The influence of the number of samples
used for the tests is also verified. For this, 50, 100 and 150 samples are used for each
load configuration. The tests are performed per device (4 electric loads) and per class
(24=16 experimental platform operating configurations). It is verified that both
sampling rate and number of samples are influenced the performance of the
classifiers, opening up possibilities for the development of new works that aim to find
optimal configurations involving such parameters. The results obtained for similar
electrical loads reached 85.94% of success when identifying a connected device and
73.75% when identifying an arrangement configuration. On the other hand, results
with non-similar electrical loads show the compatibility with the results found in the
literature, that is, varying between 92.69% and 100% accuracy.
