Research Information System
VI-INTERNATIONAL EUROPEAN CONFERENCE ON INTERDISCIPLINARY SCIENTIFIC RESEARCH, Bucuresti, Romania, 26 - 27 August 2022, pp.70-79, (Full Text)
The use of nonlinear loads based on power electronics more and more increasingly has caused
harmonic components with a significant power quality problem to increase in power systems.
Harmonic components basically cause energy losses to increase and the devices used in the
networks not to work properly or break down. Traditional passive filters and modern active
power filters are widely used in order to prevent the damages caused by harmonics. In order for
these filters to fully function, the amplitudes and frequencies of the harmonic components must
be estimated accurately, easily, and inexpensively. For the detection of these undesirable
harmonics in power electronics, artificial neural network-based techniques are increasingly
being used. In this study, a model for prediction of harmonics is developed using bidirectional
long-short term memory (Bi-LSTM), one of the unsupervised machine learning techniques.
Fast Fourier Transform was utilized to generate the dataset contains 4600 data with harmonics.
3220 of this data are divided as training data, and 1380 are divided as test data. The dataset
were preprocessed for normalization before it were given to the input of the Bi-LSTM network.
The generated dataset are used to both train and test the deep Bi-LSTM network. The proposed
approach achieved %97.3 success rate.