Abstract
A power transformer is amongst the more expensive and critical equipment installed in the power system. Unplanned power outages as a result of transformer failure have high recovery costs, reduce the life expectancy of the equipment and interrupt continuous power supply to customers. The fault occurrence in an oil-immersed power transformer results in decomposition of mineral oil which in turn causes dissolved gases to be released. To ensure reliability and availability of power transformers, mineral oil needs to be continuous ly monitored and evaluated through condition monitoring. Dissolved Gas Analysis (DGA) was developed to measure, detect, interpret and analyse dissolved gases in mineral oil of a power transformer. Condition monitoring of transformers constitutes an essential step towards prevention of unplanned breakdowns. The literature reviewed identified limitations in DGA techniques as not all incipient fault conditions can be detected by the diagnostic techniques, which makes it difficult to develop failure probability for power transformers. Considerable efforts in research have been made, in the transformer asset management field, to develop accurate models that will provide reliable incipient fault diagnosis and predict planned outage of a power transformer based on the health condition. Studies demonstrated that the Computational Intelligence (CI) and Artificial Intelligence (AI) techniques have the ability to overcome the limitations in DGA techniques. Since there is a lack of common frameworks to determine the accurate health condition and to predict the downtime of a power transformer, this dissertation provides a critical review on mineral oil sampling processes, DGA, CI and AI techniques used to improve the fault diagnostic of a power transformer and predict the planned outage timely. Preventive maintenance is outlined by the preventive model that is built using a combination of Artificial Neural Network (ANN) with DGA techniques to detect accurate incipient fault conditions. The model uses dissolved gases in mineral oil to detect to identify incipient fault conditions in a power transformer. The ANN Multilayer Perceptron (MLP) Feedforward with Back-Propagation (BP) was built using the concentrations of key combustible dissolved gases as input layer, trained using Levenberg-Marquardt (LM) algorithm to obtain the incipient fault conditions as outputs that are enlisted in gas ratio or IEEE C57.108-2004 methods. The results obtained from comparative diagnosis presented in this work show clear improvement and accuracy in the diagnosis of transformer using a combination of ANN with Rodgers ratio or IEEE C57.108-2004 methods over using the diagnostic techniques independently. Limitations of the gas ratio and Duval Triangle methods make it difficult for the model to make decisions if data presented does not fall within the ratio range scheme and fault zones of a triangle. Unidentified diagnosis can have a severe impact on the life of the power transformers. Furthermore, CI and AI algorithms such as fuzzy logic (FL), adaptive neuro-fuzzy inference system (ANFIS), machine learning (ML), ANN and non-linear autoregressive neural networks have been used in various studies to overcome the limitat ions of DGA techniques. Also, the CI and AI algorithms have their limitations. Therefore, when building the preventive and predictive models for power transformers, it is important to select the combinations of methods or techniques that will circumvent most of the limitations and provide reliable and accurate outcomes. Another model that was developed for predictive maintenance is a nonlinear autoregressive exogenous model (NARX) neural network combined with IEEE C57.108.2004. The predictive model is used to study the historical data of dissolved gases and predict the future gas levels that will be used to identify the incipient fault present and predict the planned outage based on results. The proposed models in this dissertation will manage a service life of power transformers efficiently by associating gas concentration values with incipient fault conditions and serve as an early warning in the power system network by predicting planned outage of a transformer.
M.Tech. (Electrical and Electronic Engineering Technology)