Peimankar, Abdolrahman2017-11-132017-11-132017http://hdl.handle.net/10092/14613http://dx.doi.org/10.26021/3115This thesis begins by providing an introduction to different transformer failures and the most effective condition monitoring techniques. Different failures are introduced and their corresponding fault diagnosis methods are listed to have a better understanding of failure modes and their consequence effects. An investigation into monitoring major failures of transformers using dissolved gas analysis is then presented. Various conventional, dissolved gas analysis based, fault diagnosis techniques are presented and the drawbacks of these methods are discussed. Intelligent fault diagnosis methods are introduced to overcome the problems of the conventional techniques. An overview of statistical and machine learning algorithms applied in this research is also described. Preliminary research results on transformer load tap changers fault classification are reported. A hierarchical fault diagnosis algorithm for transformer load tap changers using support vector machines is used, in which, for each fault class, a unique single support vector machine algorithm is employed. However, while the developed algorithm is reasonably accurate, the shortcomings of applying single learning algorithms are discussed and a proposal for developing a more robust and generalised transformers condition assessment algorithm is made. An intelligent power transformer fault diagnosis algorithm is then developed to classify faults of transformers. The proposed fault diagnosis algorithm is an ensemble-based approach which uses different statistical and machine learning algorithms. In the first phase of the proposed algorithm the most relevant features (dissolved gases) corresponding to each fault class are first determined. Then, selected features are used to classify transformer faults. The results of this algorithm show a significant improvement, in terms of classification. A time-series forecasting algorithm is developed to predict future values of dissolved gases in transformers. The dataset for this algorithm was collected from a transformer for a period of six months which consisted of seven dissolved gases, a loading history, and three measured, ambient, oil, and winding, temperatures of transformer. The correlation coefficients between these 11 time series are then calculated and a nonlinear principle component analysis is used to extract an effective time series from highly correlated variables. The proposed multi-objective evolutionary time series forecasting algorithm selects the most accurate and diverse group of forecasting methods among various implemented time series forecasting algorithms. The proposed method is also compared with other conventional time series forecasting algorithms and the results show the improvements over the different forecasting horizons.enAll Right ReservedTheses / Dissertations