**1. Introduction**

It is well documented that the power supply system is a composite network comprising several customers, i.e., residential, industrial, etc., operating at distinct voltage levels [1,2]. This distinct voltage level is facilitated using power and distribution transformers [2]. Consequently, the high degree of operational reliability and lucrative operation of the electrical transformers and, thereby, the power system are of considerable engineering significance. It is recognized that the performance and planned life span of a transformer is assignable to the decomposition of the dielectric system [3]. Moreover, this chain of events requires efficacious condition-monitoring procedures and diagnostic tools. For this reason, condition monitoring of dielectric oil and cellulosic paper insulation is an absolute priority to achieve a planned operational lifetime and to evade destructive failures of electrical transformers. Dissolved gas analysis (DGA) is a robust and generally acknowledged diagnostic tool in the transformer manufacturing industry. This is because DGA has the potential to divulge the electrical and thermal stresses predominating with transformer dielectric oil and cellulosic paper insulation. Prospective examination approaches for DGA fault diagnosis are reported in [4]. By and

**Citation:** Thango, B.A. Dissolved Gas Analysis and Application of Artificial Intelligence Technique for Fault Diagnosis in Power Transformers: A South African Case Study. *Energies* **2022**, *15*, 9030. https://doi.org/ 10.3390/en15239030

Academic Editors: José Matas, Saad Motahhir, Najib El Ouanjli and Mustapha Errouha

Received: 14 October 2022 Accepted: 24 November 2022 Published: 29 November 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the author. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

large, DGA fault diagnostic methods demonstrate high equivocation in examining the fault gases, given the nonlinear comportment of the fault gases and various problems with DGA fault diagnosis methods. The production of dissolved gases has a nonsequential connection with transformer duration of operation and insulation aging indicators, i.e., interfacial tension, furan, acidity, etc. [5]. This nonsequential comportment gives rise to intricacy in fault recognition when artificial intelligence algorithms are employed.

A handful of investigators have employed diverse intelligent methods comprising artificial neural networks (ANNs), fuzzy logic (FL), decision trees, etc. for diagnosing DGA faults [6,7]. As a result of the dubious accuracy and high equivocation in the classification of DGA faults, modern computational methods have been reported in recent years. These methods comprise machine learning algorithms and optimization algorithms. It is recognized that multidimensional and astronomical data samples are necessitated for training these algorithms to be efficacious. A handful of researchers have adopted these algorithms for diagnosing transformer DGA faults [8–10].

The current research contribution—This work presents a detailed investigation of transformer fault identification. Various open issues and research challenges in transformer fault identification using classical methods have been highlighted. The contributions of this research work are indicated as follows.


The novelty of the current research—The fundamental purpose of this study is to ascertain a reliable transformer fault diagnosis approach using DGA and the application of the artificial intelligence technique for fault diagnosis in power transformers. Notwithstanding that numerous research workers have worked on the application of AI to diagnose transformer faults, as shown in Table 1, very seldom has research been published on fault diagnoses using BCSVM, particularly in transformer condition assessment. A reliable fault classification and prediction algorithm are crucial criteria for developing an efficient fault identification system. Three methods are considered a benchmark of the proposed method, and it was found that the IECGR method yields a "not detectable" response to some of the set of studied case studies and ANN yields a higher fault class than the actual fault in some oil samples due to model overfitting. However after the application of the proposed BCSVM algorithm, the "not detectable" samples were effectively identified.

Many works compared the classical DGA methods in their investigations. Though there is similar work, in the current investigation, various machine learning (ML) algorithms, i.e., linear SVM, quadratic SVM, cubic SVM, fine Gaussian SVM, medium Gaussian SVM, and coarse Gaussian SVM, are compared in terms of accuracy (%), prediction Speed (objects/sec) and training time (sec). In the current study, the effect of the dissolved gas concentration levels is studied exclusively. This parameter is adopted for examining the performance of ML algorithms and two different diagnostic methods. From this investigation, it may be concluded that the proposed BCSVM algorithm is reliable in accurately diagnosing transformer faults using dissolved gas concentration levels in parts per million (ppm).


**Table 1.** Summary of applicable works.

The manuscript organization—This research has been structured as follows. Section 2 presents the fundamental principle of the SVM algorithm and proposed BCSVM. Section 3 presents the results and discussion of the proposed algorithm for fault diagnosis of transformers. A corroboration of the proposed algorithm with the real sample datasets from local transformer companies in South Africa is presented. Lastly, Section 4 provides a conclusion of the manuscript.
