Abstract:
The proposed work aims at the accurate detection
and classification of various single and multiple power quality
(PQ) disturbances. To this end, a modified optimal fast discrete
Stockwell transform (ST) with random forests (RF) classifierbased
PQ detection framework has been proposed in this paper.
In modified ST, a single signal-dependent window is introduced,
with optimally selected window parameters via energy concentration
maximization based constraint optimization. As a result of
which accurate time-frequency localization of various PQ events
is achieved, with sharper energy concentration. In classification
stage, the proposed PQ framework utilizes the RF-based classifier,
which follows the bagging approach by random selection of
features and data points, at each node, to train the classifier.
Decision stumps are used as weak classifiers, and using a simple
majority voting of these decision stumps, RF builds a strong
classifier. The RF gives less variance and less bias estimation
due to injection of randomness into the training phase, and
its performance is found to be reasonably immune to input
parameter selection. As a result of this, the classification results
of the proposed PQ framework are found to be very accurate
and quite insensitive to the presence of noise in the data. Various
test cases are presented in this paper to clearly demonstrate the
superiority of the proposed scheme. The proposed approach has
also been tested on real field data and very promising results
have been obtained.
