Intelligent Approach to Interpret Incipient Faults of Power Transformer from DGA Database

Abstract

Power transformer is most powerful and expensive tool in power system for transmitting and distributing electrical energy to all consumers. High-voltage transformers in power system are oil-immersed type transformer. Use of oil provides much needed cooling, insulation, and reduces vibrations to power transformer. Oil of the power transformer is monitored and diagnosed on a regular basis to preserve its dependability and efficiency. Dissolved gas analysis (DGA) is effective and efficient tools to interpret incipient faults. In DGA method, dissolved gases like H$_2$, CH$_4$, C$_2$H$_4$, C$_2$H$_6$, C$_2$H$_2$ are extracted from oil. Based on the gases threshold values in oil, different faults are identified. The current article focus on three traditional fault diagnostic methods IEC, Roger ratio, and Duval triangle and one artificial neural network-based intelligent method. Result spot light that intelligent methods gives higher accuracy and consistency to identify the incipient faults of power transformer while traditional methods are proved inadequate, inaccurate and inconsistent.

Introduction

Power transformer in the power system is backbone of transmission and distribution system. Performance of the power system is depend on power transformer. So its resilience affects not only on electrical energy but also operational economy [1, 2]. As a result, timely maintenance based on observed incipient faults is necessary. Nowadays most of the power transformers used in power systems are oil immerse type. And majority of the incipient faults are caused by electrical, mechanical and chemical stress; hence, oil decomposition and few gases like “Hydrogen (H$_2$), methane (CH$_4$), ethane (C$_2$H$_6$), ethylene (C$_2$H$_4$), acetylene (C$_2$H$_2$), carbon monoxide (CO), and carbon dioxide (CO$_2$)” [3] are dissolved in oil, and their threshold limits are measured in parts per million (ppm) [4, 5]. Dissolved gas analysis is the most effective method for finding early flaws (DGA) [6, 7]. Internal problems in a power transformer emit various gases that can be used to diagnose the malfunction. When many anomalies are present, the analysis is not always straightforward. Listed below are several flaws [8].

• 1. Partial discharge (PD)

• 2. Sparking discharge (SD)

• 3. Arcing discharge (AD)

• 4. Low-temperature overheating (LTO)

• 5. Middle-temperature overheating (MTO)

• 6. High-temperature overheating (HTO)

• 7. Thermal and electric faults (DT)

In partial discharge faults [9, 10], temperature has little bearing; the bulk of gases are H$_2$ and CH$_4$, with minor residues of C$_2$H$_2$, resulting in pinholes and carbonized tiny punctures in paper. Surface tracking of paper or the development of tiny amounts of carbon particles in oil is both caused by the Sparking Discharge (low-energy arcing) fault [11]. Production of large amount of H$_2$ and C$_2$H$_2$ gases is evidence of that. Further discharge may lead to arcing (high-energy discharge) type of fault which damage insulating paper. Other faults are low-, medium- and high-temperature overheating faults which may occurs at below 3000 $^\circ$C, in between 3000 and 7000 $^\circ$C, more than 7000 $^\circ$C. As an impact of these faults, larger amount of C$_2$H$_4$ gases with small traces of CH$_4$, C$_2$H$_6$ are generated and decomposed in oil. In the last stage fault identification, the DGA data will be used to analyze the transformer conditions [12]. The international standards IEC 60599 and IEEE C57.104 [5] include many diagnostic approaches for DGA [13], such as the key gas method, Dornenburg, the Roger ratio method, IEC (International Electro Technical Commission) ratio method and the Duval triangle method. The current article represents three traditional methods Roger ratio method (RRM), IEC ratio method (IRM) and Duval triangle method (DTM) and one intelligent ANN-based method are implemented on MATLAB, and analysis is performed on empirical dataset [14, 15].

Methodology

This paper represents two approaches of DGA dataset interpretation methods. In traditional DGA dataset interpretation approach, three methods [16], Roger ratio (RRM), IEC ratio (IRM) and Duval triangle (DTM) are testes over different 440 empirical observations [14, 15] with six fault categories listed above. MATLAB code is developed to test methods.

In intelligent approach, two ANN-based training functions, Levenberg–Marquardt and Bayesian regularization is developed in MATLAB. Initially, both functions were tested with same observation but due to the lower size the observations, accuracy is compromised. To achieve the better accuracy, samples sizes are increased using bootstrapping.

Traditional DGA Interpretation Methods

IEC, IEEE C57.104 and other many standard [17] have been highlighted for the incipient fault detection of the power transformer.

Roger’s Ratio Method (RRM)

Four gas ratios (R1 = CH$_4$/H$_2$, R2 = C$_2$H$_6$/CH$_4$, R3 = C$_2$H$_4$/C$_2$H$_6$, R4 = C$_2$H$_2$/C$_2$H$_4$) are used in this method to predict 11 incipient faults [18]. But after revision in IEEE C57.104-1991, R2 no longer exists [19]. As a result, only six types of incipient faults (PD, AD, SD, LTO, MTO, HTO) excluding no faults (NF) condition can be identified. Faults are recognized using ratio range scheme shown in Table 1.

RRM was developed in MATLAB and tested on empirical dataset of 440 observations; results highlight that algorithm gives the moderate result with accuracy of 63.42%. Result also enlightens that this methods fails to identify multiple faults.

Table 1 Gas analysis by RRM [16]

IEC Ratio Method (IRM)

This method is exactly identical to RRM. Three gas ratios (R1= CH$_4$/H$_2$, R3=C$_2$H$_4$/C$_2$H$_6$, R4=C$_2$H$_2$/C$_2$H$_4$) are taken into consideration [20]. Finding which region of fault is closest to the original ratio’s data point is the final step in determining the fault type. This method is able to diagnosis various overheating faults, electrical energy discharge faults and also gives the information about the normal aging (Table 2).

Table 2 Gas analysis by IRM [20]

MATLAB code is implemented and test on empirical dataset of 440 observations [14, 15], results spot that this method gives the reasonable accuracy 71.66% but in some cases this method fails to identify the faults accurately.

Duval Triangle Method (DTM)

DTM [21] uses the three gases, methane (CH4), ethylene (C2H4), and acetylene (C2H2) proportionate concentration to identify the different types of fault. Different fault types and its zone are specified in Fig. 1 and Table 4. Table 3 indicates that the normal limits and its normal rising rate of the gases are from 10 to 50% per month. Once the fault exists, this method uses percentages of %CH$_4$, %C$_2$H$_4$ and %C$_2$H$_2$ to find the exact fault category of the fault (Table 4).

Table 3 Normal limits of oil [18]

Table 4 Fault zone identification

Fig. 1

Duval triangle [21]

This traditional method gives best results among all other methods with the accuracy of approximately 88%. Later stage author [22] highlights the modified version of Duval triangle in which numerical method was employed from graph.

Other traditional algorithms like Doernenburg ratio method, key gas method [23], etc. are also used to predict the various fault, but these approaches exclusively depend upon human experience and expertise. Moreover these methods are unable to identify the multiple faults exist in the transformer oil.

ANN-Based DGA Interpretation Method

For any nonlinear input–output patterns, artificial neural network (ANN) is greatest tool to find the hidden patterns between input–output [24]. In this article, basically two training functions (Levenberg–Marquardt and Bayesian regularization) are implemented

Levenberg–Marquardt (LM) Training Function

This approach employs statistics to reduce a nonlinear function over a set of parameters. Constraint optimization challenges include nonlinear programming and least squares curve fitting. [3]. The LMA links Gauss–Newton method with (GNA) [25] and the gradient descent (GD). LMA targets second-order training speed without computing the Hessian matrix (Fig. 2).

$$\begin{aligned} P_{k+1}=P_{k}-[J^{T}J + \mu I]^{-1}J^{T}e \end{aligned}$$

(1)

Fig. 2

Flow diagram of LMA

Bayesian Regularization (BR) Training Function

Bayesian regulatory back propagation (BRP) [26] updates weight and bias variables using Levenberg–Marquardt optimization (LMO) [27]. To construct a good network, it minimizes squared errors and weights (Fig. 3).

Fig. 3

Flow diagram of BR

Figure 4 shows the basic flow diagram of the ANN-based approach [28], these can be further divided into data acquisition, pre-processing, feature selection, Training and Testing with sample data.

Fig. 4

Basic flow diagram of ANN-based training functions

Results and Discussion

Figure 5 shows the results of the three traditional methods. Total observation of 440 observations [14, 15] with different category of faults were taken for these diagnosis. All methods are able to identify the majority of the incipient faults with reasonable accuracy and Consistency. Out of all three conventional, RRM provides 65% accuracy, IRM gives 75% of accuracy and DTM yield best results and greater accuracy around 88%. Contrarily the results also shed a light that these methods are not able to identify multiple faults like discharge and overheating faults (DT).

Fig. 5

Result comparison of traditional methods

When ANN is trained with LMA, BR and BFGS quasi-Newton training functions for same 420 observations [14] [15], accuracy during training fall in between 60 and 95% and accuracy during testing (20 observations) was 50–80%, respectively. These might be occur due to following reasons.

• Less number of observations

• Uneven faults cases

• Selection of hidden layer neurons and its activation functions

So more observation were created using bootstrapping [29].

LMA training algorithm details

• Samples for training: 1195

• Samples for testing: 244

• Inputs: 05 (H$_2$, CH$_4$, C$_2$H$_6$, C$_2$H$_4$, C$_2$H$_2$)

• Output: 08 (PD, SD, AD, HTO, MTO, LTO, DT, NF)

• Hidden layer neurons: 20

• Training algorithm: Levenberg–Marquardt

• Activation function: Hyperbolic Tangent

BR training algorithm details

• Samples for training: 1195

• Samples for testing: 244

• Inputs: 05 (H$_2$, CH$_4$, C$_2$H$_6$, C$_2$H$_4$, C$_2$H$_2$)

• Output: 08 (PD, SD, AD, HTO, MTO, LTO, DT, NF)

• Hidden layer neurons: 20

• Training algorithm: Bayesian regularization

• Activation function: hyperbolic tangent

BFGS quasi-Newton training algorithm details

• Samples for training: 1195

• Samples for testing: 244

• Inputs: 05 (H$_2$, CH$_4$, C$_2$H$_6$, C$_2$H$_4$, C$_2$H$_2$)

• Output: 08 (PD, SD, AD, HTO, MTO, LTO, DT, NF)

• Hidden layer neurons: 20

• Training algorithm: BFGS quasi-Newton

• Activation function: hyperbolic tangent

Moreover, when ANN is trained using different other training functions like one-step secant, conjugate gradient with Beale Powell series, gradient descent with momentum and resilient backpropagation, the overall accuracy which will get arround 20, 33, 47, and 59.4%, respectively (Figs. 6, 7).

Fig. 6

Regression curve for BR training

Fig. 7

Mean square error for BFGS quasi-Newton training

Table 5 shows the results of three ANN training algorithm results and it highlights that with larger dataset of 1439 observations which was derived through bootstrapping [29] with different fault categories, accuracy of both training function (LM and BR) are more than 90% during training and testing phases (Figs. 8, 9 and 10). But accuracy using BFGS quasi-Newton training algorithm is poor or closer to tradtional methods. Moreover, LM and BR training functions provide best results to interpret multiple faults at same time due to overheating and discharge (DT).

Table 5 Result (Accuracy) comparison between LM, BR and BFGS quasi-Newton function

Fig. 8

Regression curve for LMA training

Fig. 9

Mean square error for LMA training

Fig. 10

Regression curve for BR training

Conclusion

In current paper, three traditional methods, i.e., Roger ratio method, IEC ratio method and Duval triangle method, were developed and tested with empirical dataset of 440 observations. Result revealed that among three methods, Duval triangle gives best fault diagnosis with considerably higher accuracy and consistency. Ratio approaches have a flaw in that they don’t cover all data regions, and occasionally ratios aren’t fit for tables. When it comes to dissolved-gas measurements, there is always some level of error. Gas concentrations and other analytical computations are all affected by this inconsistency. Furthermore, an ANN-based strategy was constructed, tested, and assessed using two distinct training functions: Levenberg–Marquardt and Bayesian regularization. Result of both training functions revealed that highest accuracy and consistency were achieved during training and testing phases, results also pined that ANN-based method easily identifies the multiple incipient faults presents in transformer. Moreover in future, the accuracy and consistency may increase by applying fusion of AI and ML techniques like (ANN + SVM), (ANFIS + SVM), (ANN + DT), etc.