Classification of power quality events using machine learning methods

Industrial plants and residential areas need to utilize electrical energy effectively. For this purpose smart grids were performed within power system where voltage and current signals are processed and monitored in detail. Thus controller systems provide solutions that keep the grid sustainability in both faulty and normal conditions. Today's industrial environment is smarter than ever before whilst Industry 4.0 is being globally approved. Most production lines include electrical devices which are able to communicate each other and controlled from a single station with automation systems. Most of those elements have an internet connection link known as industrial internet. Development of smart technology with industrial internet comes with a need of monitoring. Monitoring technologies are emergent systems that focus on fault detection, grid self – healings and online tracking of power quality issues. As a comprehensive and complex system, electrical grid includes numerous components in control center, switchyards, transmissions and distribution modules. In a complex system, monitoring process has a great role. Smart measurement units collect all required data signals roaming in electrical grid which are especially voltage and current signals to monitor power quality issues. Using the collected data, all the required information to the operator is provided to specify a diagnosis and prevention abnormal operation. Present study deals with one of the essential part of an electricity grid monitoring system called power quality event classification in a manner of machine learning topic. In this study, an intelligent pattern recognition system which performs classification of power quality events has been designed. The backbone of the study which is the dataset has been gathered from the substation centers all over the Turkey. In the feature extracting stage which builds a meaningfull whole from the raw dataset, Histogram, Permutation Entropy, Local Peaks, and instantaneous time domain based techniques have been used. In addition to those techniques of feature extracting which are novel to power quality event classification field, commonly used Discrete Wavelet Transform features are determined to construct a robust feature set. In decision stage, a machine learning based structure which is Extreme Learning machine and its enhanced versions Weighted Extreme Learning Machine and Sparse Bayesian Extreme Learning machine have been designed. Extreme learning machine and its enhanced versions are preferred in most of the research fields because of their learning structure which does not include any iterative process contrary of \emph{back propagation learning} based methods. Findings of thesis are analyzed in detail and a comprehensive evaluation method is used with various performance criteria. All the analyzes are held in MATLAB environment. The algorithm of the best performance is proposed to build a generic product as an output for this thesis.