Abstract:
Power quality has become a major concern recently because of increasing number
of sensitive loads being connected to the power system. Degradation in power qual-
ity is normally caused by power-line disturbances, malfunctions, instabilities, short
lifetime, failure of electrical equipments etc. In order to improve power quality, the
sources and causes of power quality (PQ) disturbance events must be known apri-
ori to take appropriate mitigating actions. However, to determine the causes and
sources of PQ disturbances, it is important to detect, localize and classify them.
For the classi cation of PQ disturbances, a wide range of signal processing meth-
ods have been reported in the literature. Since, PQ disturbance is a non-stationary
signal, development of a PQ disturbances classi cation method, which is simple yet
e ective in handling practical conditions, such as multiclass PQ disturbances, ran-
dom selection of increased training and testing dataset, and presence of noise, is
still a challenging task. In this thesis, a new method for the classi cation of PQ
disturbances exploiting higher order statistics in the Empirical mode decomposi-
tion (EMD) domain is proposed. A PQ disturbed signal is rst analyzed in terms
of intrinsic mode functions (IMFs) by using EMD operation. The Higher Order
Statistics (HOS), such as variance, skewness and kurtosis of the rst three extracted
IMFs are then utilized to form the feature vector. The feature vector thus obtained
when fed to the Probabilistic Neural Network (PNN) and k-Nearest Neighborhood
(kNN) classi ers separately is found to be capable of classifying the multiclass PQ
disturbance signals even in the presence of noise. Moreover, as expected, the clas-
si cation accuracy is found to be enhanced using the proposed feature set while
increasing the training and testing dataset. For the characterization of PQ dis-
turbance signals, mathematical models of eleven classes of disturbances are used.
Simulations are carried out to evaluate the performance of the proposed method in
terms of e ciency derived from the confusion matrix and CPU time representing
the computational burden. It is shown that the proposed method outperforms some
of the state-of-the-art methods with superior e cacy in stringent conditions.
