Machine learning for power transformer diagnostics

Due to the significant human intervention required, it is particularly challenging to collect adequate quantities of suitable DGA data, which contain a range and breadth of transformer conditions and can be easily labelled with known issues and emergent or established faults. <br><br>A new and hybrid method for transformer DGA condition monitoring using unlabelled datasets is proposed in this thesis. The work required construction of a large-scale fault database for the proposed ML method.

Moreover, performance evaluation and assessment of common AI models, including artificial neural network (ANN), support vector machine (SVM), K-Nearest Neighbour (KNN) and Decision tree using the constructed database is presented.<br><br>Obvious differences in diagnostic assessment of transformer faults are a result of unequal quantities of datasets representative of the range of fault types. Fault classification using a large quantity of training data is better than comparable assessment using small quantities.

In resolving problems with unbalanced datasets, this work investigated and proposed a classification method using a Gaussian process (GP). Comparative results between the proposed model and classical AI models are presented in this thesis to demonstrate effectiveness and improvements in diagnostic performance using GP multi-classification (GPMC), especially for small-quantity datasets. The reliability of the proposed ML method in comparison to conventional DGA approaches is also presented.<br><br>This thesis presents works investigating a hybrid AI model to resolve the difficulties with imbalanced datasets.

Various data augmentation techniques are employed for data pre-processing, which balances the volume of training data in each fault. AI models then render diagnostic fault determination using processed datasets. An evaluation metric is used to analyse and compare the performance of different models.

Experimental results demonstrate benefits for minority classes such as rare but significant transformer faults, including partial discharge (PD) and combination faults (DT), rendering improved diagnostic performance.<br><br>To forecast potential transformer faults, in addition to diagnosis at the instant a fault has occurred, a dynamic fault prediction model based on long short-term memory (LSTM) is proposed in this thesis.

Time sequence information within DGA datasets is used to predict dissolved gas concentrations and the reliability of the proposed LSTM method is presented in comparison to common regression approaches in terms of error metrics. Additionally, different AI models are used to forecast future transformer health conditions using predicted results from LSTM outcomes, with comparative results presented.<br><br><br>