Classification of power quality problems in power systems using ai methods

In the operation of power systems, it is aimed that the energy reaching the consumer is of high quality, safe and uninterrupted. One of the most important reasons for the increase in energy demand and the problems that arise with it is the stability problem of the system. Increasing energy demand, especially with the increase in population and industrialization, brings power systems to be operated under more severe conditions. This causes the lines to be overloaded and congestion in the transmission lines. Due to the high installation costs of new transmission lines, the systems are operated at their maximum capacity and stability limits. In the thesis, the factors affecting voltage stability and power quality problems, which is one of these factors, are emphasized. Defining the problem is critical for its resolution. For this reason, signals that disturb the voltage stability in the network are modeled by using mathematical equations in MATLAB environment. By classifying the generated signals, it is aimed to detect the decay signals occurring in the network. Since the malfunction in the system can be detected with the created model, it will be easier to take steps to ensure the stability of the system. In the classification process, both machine learning methods such as K-Nearest Neighborhood Algorithm (KNN), Support Vector Machines (SVM), decision trees and Artificial Neural Networks (ANNs) were used. Fourier Transform (FT), S-Transform (ST), Hilbert Huang Transform (HHD), Wavelet Transform (WT) and Discrete Wavelet Transform (DWT, Discrete Wavelet Transform) are widely used for feature extraction. In this study, 13 features such as energy, standard deviation, average of absolute value, curvature, kurtosis, average (median) absolute deviation, average frequency, median frequency, total harmonic distortion, rms, peak size of rms ratio and entropy were extracted from the signals produced in this study and given to machine learning classifiers and ANNs. In addition, feature extraction was performed with the DWT method and the results were given comparatively. As a result of the study, the highest accuracy of 94.4% was found with Cubic SVM in the classification made by removing 13 features without wavelet transform.