Application of Dissolved Gas Analysis for Transformer Condition Assessment Using Computational Intelligence

Dissolved gas analysis is widely used in the detection of power transformer incipient faults. Presently, the interpretation of the dissolved gases is dependent on an expert's opinion at the power utility in the Republic of South Africa and often limited in the representation of the fault conditions. One of the drawbacks of relying on experts is the probability of error in judgement and diagnosis. Also, there are few experts in the field of transformer condition assessment. This paper aimed at evaluating the performance of the artificial neural network and support vector machine algorithms for the different established gas ratio concentrations methodologies on the Orange platform. To overcome the inaccuracies in these methods, a modified Doernenburg gas ratio method was proposed. Testing the model with a credible dataset of 922 field cases shows that the artificial neural network performed better with the accuracy of 97.5% of correctly predicted faults while the support vector machine is 90.4%. The modified method enhances the accuracy of the decisions in fault identification, thereby indicating a good application prospect in engineering practice.