A novel approach for oil-based transformer fault identification in electrical secondary distribution networks

Abstract

Oil-immersed transformers are mostly installed equipment in electrical power networks. Thus, ensuring the reliability of these equipment is paramount. Fault identification is one of the measures taken to ensure the reliability. Concerning fault identification mechanisms, several methods exist; one of which is the dissolved gas analysis (DGA) tool that has been widely used purposefully for oil-immersed transformers. The tool is used to capture and analyze the gases dissolved in the oil. Studies have proposed various fault identification methods based on this tool such as the Dornenburg, key gas, and International Electrotechnical Commission (IEC) standards methods using DGA data. However, the accuracy of these methods seems to be limited by the context, such as the type and number of parameters used. In this study, the novel hybrid fault identification method based on multilayer neural networks (MLANN) and expert knowledge is proposed to improve the accuracy fault identification process by considering the specific characteristic parameters extracted from the historical DGA dataset. The proposed hybrid method improves accuracy by 14.46% when compared to 12 benchmark methods. This improvement portrays that our method can be integrated into a scheduling platform for maintenance decision support in an electrical power network with accuracy. Furthermore, the promising accuracy in transformer fault identification is expected to increase safety and power reliability.

1. Introduction

The operation state of the electrical power equipment is directly related to the safety and stability of the power system. Some power equipment is more critical than others; if an in-service failure occurs, it can lead to extended unplanned downtime and revenue loss to the utility company. Transformers are one of the most expensive and critical equipment in power systems. Following the expensive nature of such power equipment, the lack of accurate fault identification methods to enhance proper equipment operations may impact utility companies' safety and economic issues. Therefore, scholars have proposed several methods that attempt to deal with the accuracy of a fault identification process.

Extensive research has been conducted attempting to improve the accuracy of fault identification methods, especially in oil-immersed transformers [1]. Among several methods, the dissolved gas analysis (DGA) is a widely used and modern technique in assessing the condition of an oil-immersed transformer [2]. Other traditional methods such as Doernenburg, key gas, and IEC standards have been used for identifying transformer fault types [3]. In addition to the modern fault identification process, scholars established various fault identification techniques to improve the accuracy of fault identification methods using DGA data [4]. However, despite the effort taken by researchers to establish a robust technique for identifying transformer faults, their proposed methods are limited to identifying some faults only. The most common faults identified are partial discharge (PD), discharge of low energy (D1), discharge of high energy (D2), and thermal faults. This limitation arises because the authors did not consider some characteristic parameters, such as water content specific to a certain context [5]. Thus, some of the faults may likely remain unidentified. Failure to identify some of the faults associated with the transformer may lead to equipment failure and in turn, compromises power system reliability. Adhering to the importance of having a more accurate method of transformer fault identification, the authors of this work are motivated to propose an improved mechanism.

Therefore, this study proposes a hybrid method for transformer fault identification to improve accuracy by considering context-based characteristic parameters and gas ratios. With context-based introduced characteristic parameters, more fault types were identified compared to the existing methods proposed in the literature. The hybrid method proposed in this work seemed to identify the single-characteristic parameter's fault, such as moisture content and oil leaks, that the existing methods could not determine. The ability to identify these two fault types helps to prevent transformers from the damage caused by the moisture content and flashover. The theoretical and practical contributions of this work are as follows:

• •Establishment of a hybrid transformer fault identification method based on DGA data with improved accuracy.
• •Establishment of practical transformer fault types that can be analyzed based on DGA data. This study presented ten types of faults instead of six common fault types presented in the literature.
• •Establishment of the fault identification capability of the traditional fault identification methods.

This paper is organized as follows; the analysis of some similar works is discussed in section 2, while the material and methods used to achieve the objectives are discussed in section 3. Furthermore, results from the analysis process are presented and interpreted in section 4. Last, the concluding remarks are summarized in section 5.

2. Related work

First, the refined DGA method of the International Electrotechnical Commission (IEC) for transformer fault identification was proposed to improve the performance of the traditional IEC [6]. Unfortunately, the IEC approach is limited to the identification of some fault types; therefore, its output is marked as undetermined [7]. Moreover, the refined method uses the four types of gas ratios instead of the three ratios previously used by the traditional IEC method. However, the proposed method identifies five transformer fault types and attains an accuracy of 66.6%.

Second, the study by Dhini et al. [8] proposed a method for DGA transformer fault identification by considering the main five dissolved gases in the transformer oil. The study uses the gas-limit for fault identification which causes some interference between transformer fault types. With the proposed method, the authors achieved an accuracy of 71.5%. Due to the limitation of fault-type interference, the proposed method was modified to include seven gas ratios, and the accuracy increased to 84.71% [9]. In addition, the proposed method can accurately identify six types of transformer fault types based on DGA data.

Third, another study by Dhini et al. [8] proposed a data-driven approach for transformer fault identification based on the DGA data. The study tested different machine learning models. The SVM model was selected based on its ability to perform better with a limited number of DGA data sets for training the model. Five types of transformer faults were correctly identified in this study. Limited identified fault types are due to an inadequate number of input parameters used during the training process. The inadequate number of input parameters may leave other types of transformer faults uncategorized.

Fourth, a study by Ref. [5] proposed a new approach that enhanced the accuracy of the existing transformer fault identification methods based on DGA data. The new gas ratios and limits were introduced to improve the performance of the proposed approach. The comparative result showed that the proposed method achieved a higher accuracy (88.86%) than other DGA techniques presented in the literature. However, some significant ratios, such as CO₂/CO were not considered [10]. Due to this, the proposed method fails to identify some of the faults and faults identified as undetermined. Fewer input parameters reduce the proposed method's accuracy, especially when the ratio of CO₂/CO is below 3 parts per million (ppm). These two parameters are mostly used to evaluate the transformer's cellulose insulation degradation using DGA data [11]. The accuracy of the proposed heptagon graph method for transformer fault type identification achieved an accuracy of 89.41%.

Last, to enhance the accuracy of transformer fault identification methods, scholars include more characteristic parameters such as carbon monoxide (CO) [12]. The proposed study uses 449 DGA data collected from laboratories and literature. Results illustrate that the proposed method has a high accuracy (92.8%) compared with other transformer fault identification methods based on DGA data. Recently [13], proposed a novel method for transformer fault diagnosis based on DGA data and proved to increase recognition accuracy by 1.3% compared to the traditional IEC method. The analytical results of the related works are presented in Table 1.

Table 1.

Analytical findings from the related works.

S/N	Authors	Method	Observation
1	[14]	deep belief network	•Six fault types are identified •The accuracy needs to be improved
2	[15]	Kalman filter sensor fusion	•Three fault types are identified; Partial Discharge (PD), Arc Failure (AF), Thermal Failure (TF), and Normal Condition (NC) •Accuracy needs to be improved
3	[15]	k-means clustering	•Improved accuracy •Six fault types are identified
4	[16]	Stacked Sparse Auto-Encoders	•Six fault types are identified
5	[17]	Random forest model enhanced by synthetic minority over-sampling technique	•Seven fault types are identified •It further identifies the sub-types of thermal faults
6	[18]	Machine learning (ML)	•Higher accuracy and consistency in identifying the incipient faults of power transformer •Traditional methods are proven inadequate, inaccurate, and inconsistent •Only basic fault types are identified
7	[5]	Teaching-learning based optimization	•Five faults are identified •The accuracy needs to be improved
8	[19]	Random Forest Classifier Aided by Data Clustering Method	•Use the traditional Duval pentagon method •The accuracy needs to be improved
9	[20]	Using SVM-BA Classifier	•Six electrical and thermal fault classes were categorized
10	[7]	Fuzzy logic	•There is an improvement. It improves the accuracy

3. Material and methods

3.1. Requirement identification

This study uses 8 characteristic parameters; hydrogen (H₂), carbon monoxide (CO), methane (CH₄), ethylene (C₂H₄), water (${\mathrm{H}}_2\mathrm{O}$), acetylene (C₂H₂), ethane (C₂H₆), and CO₂. Guided by expert knowledge, water (${\mathrm{H}}_2\mathrm{O})$ was the dominant parameter that affects the transformer's life in the ESDNs, especially during the rainy season [21]. Also, an observation was made that the transformer material may add water to the oil, promoting premature transformer cellulose aging. Noting these observations from expert knowledge, this study used the eight transformer parameters for transformer condition classification and fault identification. The output parameters were categorized into ten groups namely; Partial discharge, Low-energy discharge, High-energy discharge, Thermal fault T > 300 °C, Thermal T > 300°C–700 °C, Thermal fault T > 700 °C, cellulose deterioration due to aging (CDA), Cellulose deterioration due to other factors (CDF), Contamination in a tank (CIT), Water leaks into the oil and no fault. Table 2 summarizes the input parameters used in the fault identification process.

Table 2.

Fault identification input parameters.

Method	Input parameters
IEC ratio	C2H2/C2H4, CH4/H2, C2H4/C2H6
Gas ratios	CO2/CO, C2H2/H2
Gas limit	CO, CO2, ${\mathrm{H}}_2\mathrm{O}$

3.2. Data collection

A total of one thousand DGA historical datasets with corresponding faults were collected from Tanzania Electric Supply Company (TANESCO). The sample data were divided into two parts, 800 training samples and 200 testing samples. The gases and gas rations were used as inputs and fault types were used as the output.

The gas limits and gas ratio limits used to test the faults are presented in Table 3, Table 4, and Table 5.

Table 3.

Gas limits with corresponding fault type.

Gas concentration			Fault Type
CO2	CO	${\mathrm{H}}_2\mathrm{O}$
>2500	>350	>20	Cellulose aging
<2500	<350	>20	Leaks in oil

Table 4.

Gas-ratio limits with corresponding fault type.

Fault type	Description	CO2/CO	C2H2/H2
CDA	Paper deterioration due to aging	<3	NS
CDF	Cellulose deterioration due to other factors (stress faults)	>11	NS
CIT	Contamination in tank	NS	2–3

NS: Not Specified.

Table 5.

IEC-60599 gas ratios.

Fault type	Description	C2H2/C2H4	CH4/H2	C2H4/C2H6
PD	Partial discharge	NS	<0.1	<0.2
D1	Low energy discharge	>1	0.1–0.5	>1
D2	High-energy discharge	0.6–2.5	0.1–1	>2
T1	Thermal fault T> $300°C$	NS	NS	<1
T2	Thermal T> $300°C-700°C$	<0.1	>1	1–4
T3	Thermal fault T> $700°C$	<0.2	>1	>4

NS: Not Significant.

3.3. Data preprocessing

The hybrid method starts by loading the pre-processed data for training purposes. The min-max normalization method was used to normalize the acquired DGA historical data for easy manipulation. The outliers were removed from the dataset to smooth the normalization process.

3.4. Model training

The Multi-Layer Artificial Neural Network (MLANN) was trained for transformer condition classification and further, the expert system was integrated into MLANN to compute the fault types.

The MLANN structure consisted of the input, hidden, and output layers. In this study, the input layer of MLANN contained ten input parameters, four hidden layers, one output layer, and a logistic activation function (Fig. 1). The output provides information regarding the transformer fault types.

Fig. 1.

Structure of MLANN fault identification based on DGA data.

The activation function used in designing the MLANN is given by a sigmoid function represented in Equation (1), and the important components of MLANN are presented in Table 6 [22].

$$“f\left(x\right)=\frac{1}{{1+e}^{-x}}”$$ (1)

Table 6.

Important components of MLANN.

Component	Value
Optimizer	Adam
Layer	4
Neuron	35
Epochs	1500
Batch size	40
Activation function	Logistic

The inputs in Equation (1) include single gas for the classification of transformer condition and gas ratios for fault identification. The gas ratios and gas-ratio limits are presented in Equation (2) and Table 3 respectively.

$$\left(\begin{array}{c}r1\\ r2\\ r3\\ r4\\ r5\end{array}\right)=\left(\begin{array}{c}{C}_2{H}_2/C_2{H}_4\\ C{H}_4/H_2\\ C_2{H}_4/C_2{H}_6\\ C{0}_2/C0\\ C_2{H}_2/H_2\end{array}\right)$$ (2)

3.5. Identification of fault types

The procedures for identifying the fault type stepwise, as presented in the flowchart (Fig. 2), are as follows:

Step 1

The two gas ratios, (r1) and (r2) are examined to check that they meet the conditions, condition1 (cond1) and condition2 (cond2) for the PD fault type. If the requirements for PD are satisfied, then the test for other fault types will check the multiple faults. If the condition is not met for other faults, then it is concluded that there is a PD fault type.

Step 2

If the requirements for PD are not satisfied, the conditions (cond1, cond2, and cond3) for D1 are examined through testing r1, r2, and r3. If the requirements are met, it is concluded that there is a D1 fault type; if not, the process proceeds to step 3.

Step 3

The conditions (cond1, cond2, and cond3) for D2 are tested based on the limit of gas ratios (r1, r2 and r3), as presented in equation (2). If all three conditions are satisfied, it is concluded that the D2 fault type occurred. The thermal fault is tested in the next step if the conditions are not met.

Step 4

The thermal faults are categorized into three, depending on the temperature. The T1 fault type is tested by examining conditions (cond1 and cond2) for T1. If the conditions are met, it is concluded that the T1 fault type occurred; otherwise, further testing is done to investigate other types of thermal faults.

Step 5

The T2 fault type is examined by testing conditions (cond1, cond2 and cond3). The gas ratio limits for T2 are tested, and if all conditions are met, the occurrence of the T2 fault type is concluded.

Step 6

In this step, the conditions for T3 are examined. The gas ratio limits for T3 are tested through conditions (cond1, cond2, and cond3). If the conditions are satisfied, it is concluded that T3 occurred.

Step 7

The fault type associated with non-combustible gases is tested in this step. The gas ratio (r4) is examined to test if the condition (cond4) for the CDF fault type is met. If the condition is not met, then the system proceeds to step further.

Step 8

Another test is performed in this stage to check the occurrence of the CIT fault type. The gas ratio (r6) is examined to see if it meets the condition (cond5). If the condition is satisfied, it is concluded the CIT occurred. Otherwise, the last step is performed.

Step 9

The examination based on the gas limit is performed to test the existence of CDA and leaks into the oil. If the condition (cond6) is met, there are leaks into the oil; otherwise, the CDA fault type occurs.

Fig. 2.

Flowchart of the hybrid fault identification approach.

4. Results and discussion

After an intensive literature review, requirement gathering from TANESCO, and data processing, it was established that the existing fault identification algorithms could effectively identify up to 6 types of faults out of 10 transformer faults identified in TANESCO. Fig. 3, Fig. 4, Fig. 5, Fig. 6, Fig. 7 present simulation results for the faults identified yearly, between 2016 and 2020, by applying three fundamental transformer fault identification methods. The simulation results indicate that each algorithm can identify a few transformer faults, suggesting inaccurate maintenance decision-making processes. Therefore, this research proposed the hybrid approach to identify a wider range of oil-immersed transformer faults. The DGA data collected from 2016 to 2020 were used to test the performance of the proposed approach. Table 6 shows that the hybrid approach can accurately identify all fault types compared to the IEC 60599 standard, gas ratio, and limit-based methods. Therefore, the proposed fault identification approach is more generalized and can be used in maintenance decision-making in electrical power networks.

Fig. 3.

Comparison of fault identification method based on 2020 DGA data.

Fig. 4.

Comparison of fault identification algorithm based on 2019 DGA data.

Fig. 5.

Comparison of fault identification algorithm based on 2018 DGA data.

Fig. 6.

Comparison of fault identification algorithm based on 2017 DGA data.

Fig. 7.

Comparison of fault identification methods based on 2016 DGA data.

For example, in 2020, 9 fault types were identified based on the actual DGA data collected from the utility company. The same data were used to test the performance of the different methods for transformer fault identification. Based on the result presented in Fig. 3, some methods failed to identify the fault types that occurred in 2020 correctly. Compared to the existing method, the hybrid approach shows a high capability of identifying the fault types based on the DGA data. The same trend has been observed in previous years, 2019 to 2016, as presented in Table 7.

Table 7.

Comparison of fault identification capability for different methods.

Year	Number of Fault Types	IEC 60599	Gas ratio	Limit-based	Hybrid
2020	9	4	3	1	9
2019	8	4	3	1	8
2018	8	5	2	1	8
2017	9	4	2	1	9
2016	7	4	2	1	7

The sample of 97 datasets was used to evaluate the proposed method's performance compared to the existing ones. The identified fault types were compared with the expected sample data from the Tanzania utility company. The result of the proposed fault identification method is shown in Table 8.

Table 8.

Accuracy of the proposed hybrid algorithm.

Fault types	Sample	PD	D1	D2	T1	T2	T3	UD	NF	Accuracy%
PD	29	16	3	5	4	1	0	0	0	55.172
D1	1	0	1	0	0	0	0	0	0	100
D2	32	2	2	22	2	2	2	0	0	68.75
T1	2	0	0	0	2	0	0	0	0	100
T2	29	0	0	2	2	22	3	0	0	75.89
T3	4	0	0	0	0	1	3	0	0	75
All	97	18	6	29	10	26	8	0	0	45.36

The accuracy was investigated for the different identification methods and presented in Table 9, Table 10, and Fig. 8. The standard methods in some conditions failed to identify the fault types correctly. The larger number of incorrect identifications decreases the accuracy. In literature, research has been conducted to increase the accuracy of the identification methods. However, it still needs improvement to be used in a practical application. Therefore, designing a hybrid identification method was justified as the method with improved accuracy than the standard methods.

Table 9.

Accuracy of eight benchmark fault identification methods.

Fault types	IEC-Mod	Rogr4-Mod	Duval	Rogers'4	IEC-60599	Clustering	Probability	CSUS
PD	41.3	34.4	3.44	3.44	31.03	10.34	17.24	13.79
D1	0	0	0	0	0	0	0	0
D2	25	6.25	25	0	0	21.87	15.62	12.5
T1	100	100	0	100	0	100	100	100
T2	27.5	27.58	20.68	3.44	31.03	10.34	27.58	0
T3	0	0	75	0	0	0	25	0
All	30.9	22.68	18.55	4.123	18.55	15.46	21.64	10.3

Table 10.

Accuracy of the other benchmark and hybrid algorithm.

Fault types	IEC 60599 Refined	Rogers'4 Refined	NPR	SVM	Hybrid
PD	37.93	10.34	20.7	24.13	55.17
D1	100	100	0	0	100
D2	3.12	6.25	18.8	18.75	68.75
T1	0	100	100	100	100
T2	31.03	6.89	37.9	41.37	75.89
T3	0	0	0	0	75
All	22.68	10.3	25.8	27.83	45.36

Fig. 8.

Comparison with the benchmark.

Based on the evaluation of the proposed fault identification method with other DGA methods proposed in the literature [5], it has been revealed that the proposed method improves the identification accuracy by 14.46%. However, despite the improvement the method has demonstrated, there are some limitations associated with it. First, the proposed method applies to some parameters of an oil-immersed transformer. This makes it a bit challenging to generalize to all kinds of transformers. Second, the method achieves better accuracy by adding more parameters to the model relative to the previous methods. Therefore, our method is also limited to those parameters used in the model.

5. Conclusion and future work

The hybrid approach based on a Multilayer Neural Network and expert knowledge has been proposed in this study. The approach was mainly intended to improve the accuracy of the existing transformer incipient fault identification. Despite the six common transformer faults, this study identified more fault types such as cellulose deterioration due to aging (CDA), Cellulose deterioration due to other factors (CDF), contamination in a tank, and water leaks into the oil, that the existing literature could not. Furthermore, the study introduced new characteristic parameters such as water (${\mathrm{H}}_2\mathrm{O}$) which can be used to identify the single-gas fault types.

When compared with the traditional and recent fault identification methods based on DGA, the proposed hybrid approach showed promising results. The results were tested and found to attain improved accuracy by 14.46%. The achieved improvement demonstrated by the proposed hybrid approach guarantees an efficient fault identification mechanism. Such efficient mechanisms may benefit the emerging predictive maintenance systems in various contexts that demand an accurate and practical fault identification method such as smart grid. Furthermore, promising improvements obtained in this study signify that the hybrid fault identification approach can be adopted in maintenance decision-making.

First, the authors of this work recommend further research to include all types of transformers to establish a generalized fault identification method. Second, future research should involve feature selection algorithms to determine the optimal number of parameters to be used in a fault identification model. Third, other studies should focus on the integration of the prediction engine into maintenance architecture to facilitate predictive maintenance.

Data availabillity

The data used in this research is private. Therefore, it can be available on request sent to the authors of this manuscript.

Funding

Not applicable (No source of funding received for this research).

CRediT authorship contribution statement

Hadija Mbembati: Formal analysis, Data curation, Conceptualization. Hussein Bakiri: Writing – review & editing, Visualization, Methodology.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Biographies

Hussein Abubakar Bakiri is a Lecturer and Researcher from the Department of Computer Science at the Institute of Finance Management, Dar es Salaam. He graduated from the University of Dar Es Salaam (UDSM) with a Bachelor's degree of Science in Computer (2006) and a Master's degree in Web Technologies (2008) from the University of Southampton (UK). He also graduated with a Ph.D. degree in Computer and IT Systems Engineering in the area of Grid automation in 2022 from the University of Dar es Salaam. Dr. Bakiri majors in exploring the quality of electricity data such as outlier detection and data cleansing, as well as forecasting load demand for power service restoration.

Hadija Mbembati is a Lecturer and Researcher at the College of Information and Communication Technologies at the University of Dar es Salaam. She graduated from the University of Dar Es Salaam (UDSM) with a Bachelor of Science in Electronics Science and Communications (2008) and a Master's degree in Electronics Engineering and Information Technology (2015). She graduated with her Ph.D. degree in Computer and IT Systems Engineering in the area of Smart Grid in 2022 at the University of Dar es Salaam. Dr. Mbembati is interested in exploring the use of Condition–Based Maintenance for failure prediction of critical electricity equipment in secondary distribution networks.

Acronyms and Abbreviations

AI

Artificial Intelligence

ANN

Artificial Neural Network

DGA

Dissolved Gas Analysis

ESDN

Electrical Secondary Distribution Network

MLANN

Mult-Layer Artificial Neural Network

TANESCO

Tanzania Electric Supply Company

UDSM

University of Dar es Salaam

SIDA

Swedish International Development Agent

CDA

Cellulose deterioration due to aging

CDF

Cellulose deterioration due to other factors

IEC

International Electrotechnical Commission

r1

Ratio 1

r2

Ratio 2

r3

Ratio 3

r4

Ratio 4

Cond1

Condition 1

Cond2

Condition 2

Cond3

Condition 3

Cond4

Condition 4

PD

Partial discharge

D1

Discharge of low-energy

D2

Discharge of high energy

T1

Thermal fault below $300°C$

T2

Thermal fault above $300°C$

T3

Thermal fault above $700°C$

DT

The mixture of heat and electrical abnormalities (DT)

UD

Undefined

NF

No fault