Abstract
In this dissertation, a machine-learning algorithm for the automatic classification of single-source partial discharge (PD) in high-voltage (HV) power transformers is proposed. The research includes the design, construction, testing and analysis of the PD classification algorithm with the use of artificial intelligence (AI) techniques. The suggested classification system includes data preprocessing, feature selection and dimensional reduction techniques together with the classifier algorithms that are applied for identification of partial discharge patterns. The data used for training the algorithm is collected from power transformers of different sizes and voltage ratings. The collected data is pre-processed, and only applicable features are applied to derive the PD patterns. Feature extraction techniques are then used to extract features for the single source PD identification. Some renowned PD classification algorithms are applied to the developed PD patterns to derive an all-inclusive classification system as presented. The developed system is then tested by supplying it with known PD data and evaluated on the accuracy of identification and classification of the test data. The test data is derived from different sources, including industry transformers of different sizes and voltage ratings as well as laboratory tests of simulated PD patterns. This allows for further testing of the accuracy and limit of the derived classification system. The testing results demonstrate that the classification system is effective in identifying and classifying single-source PD. The outcome of this dissertation allows for industrial transformer owners and users, as well as electrical utilities, to have a constant, uninterrupted partial discharge monitoring system which can be directly connected to PD analysis equipment and be able to continuously monitor the PD behaviour of their transformers. By so doing they can recognize increases in PD activity, thus deterioration of solid insulation, at an early stage. Thus, providing time to either continue with close monitoring of the PD behaviour or perform necessary repairs before continued deterioration.