Download files
Abstract
On-line partial discharge (PD) monitoring that involves long term continuous monitoring is important for condition assessment of power system equipment. Tremendous amount of PD data can be extracted from such monitoring process and stored. Due to the tremendous data size, useful information is not possible to be realized through the use of traditional techniques of data handling and analysis. To extract hidden useful information from the large database, data mining is the perfect solution. Thus, the application of data mining techniques to PD aspect for both predictive and descriptive purposes will be demonstrated in this thesis.
Experimental setups were prepared in the laboratory to produce basic PD types such as corona discharges, surface discharges and internal discharges. The PD patterns and PD analysis results were also presented in the thesis. Prior to the application of data mining techniques to PD data, several essential processes are required. Firstly, the measured PD signals need to be de-noised, followed by the separation of PD signals from multiple PD sources. A method to separate PD signals from multiple PD sources were proposed by the author. The proposed method requires the use of a simultaneous multi-channel PD measurement system together with clustering analysis and the application of the principal component analysis (PCA) techniques.
This thesis also aims to find the current state of art for PD fault type classification task in the aspect of predictive data mining application on PD. Two types of PD input methods were suggested to be applied as input to the predictive data mining techniques. The first input method is the statistical operators of the phase resolved distributions; whereas the second method is by using the raw PD data samples of the phase resolved distributions. In this study, three data mining techniques were utilized and they were namely, backpropagation neural network (BPN), self-organizing map (SOM) and support vector machine (SVM). The main objective here is to determine the combination of PD input method and predictive data mining that can achieve highest PD classification accuracy.
Thus far, descriptive data mining application on PD has yet been attempted by researchers around the world. Two descriptive data mining techniques were proposed. The first technique is based on decision tree (DT) and genetic algorithm (GA), whereas the second technique is a combination of rule extraction based on support and prototype vectors and GA. The proposed descriptive data mining techniques search for the characteristics and formulate explicit rules that identify/differentiate specific PD fault type from others. In this thesis, the use of descriptive data mining techniques as an aid to the application of predictive data mining, in order to solve the ‘black-box’ problem were also demonstrated.