Business

 Use Of AI For DGA Of Transformer Oil 

 Introduction

Power transformers are one of the most expansive and vital assets of a power utility for the transmission and distribution of electrical power systems. The smooth and trouble free functioning of power transformers is essential for power flow from generating stations and grid substation. Unplanned unavailability of a power transformer due to any fault condition causes interruptions in power flow and require expensive and long repairs. An incipient fault in a transformer due to aging of internal components of the transformer must be detected as soon as possible. This detection can initiate predictive maintenance and prevent the further deterioration of the transformer. Condition monitoring approaches can be divided into two groups i.e. on-line and off-line. Transformer ageing process of insulating oil and cellulose (Paper) materials is determined by measuring different parameters such as dissipation factor, resistivity, breakdown voltage of oil, dissolved gases concentration, acidity, moisture content, sludge content, furan analysis, interfacial tension (IFT) analysis, flash point, pour point, viscosity etc. Out of these parameters dissolved gases concentrations are measured by a technique called dissolved gas analysis (DGA). Dissolved gas analysis is one of the most popular and practical methods to identify incipient faults in a power transformer. Dissolved gas analysis is based on IEC60599 standards. The DGA consists of several fault diagnosis techniques, such as characteristic or key gas analysis method (KGM), Rogers gas ratios method (RRM), Doernenburg Ratio Method (DRM) and Duval Triangle Method (DTM) to detect the different types of faults.  DGA includes detection and estimation of concentration of gases. These gases are called characteristic gases.

Typical gases that are formed in power transformation due to incipient faults are hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), methyl ethylene (C3H6), Propane (C3H8), carbon dioxide (CO2), carbon monoxide (CO) and oxygen (O2). Each of these gases begin to form at different values of temperature and dissolve inside the transformer oil. 

The DGA data provides information about the overall internal health of the    transformer, prior alert about initiating faults or pending faults and information about faults that have occurred. This information is useful to plan predictive maintenance to prevent any future outage of the transformer.  In conventional methods as aforesaid, the result of fault diagnosis depends upon analytical expertise of humans. Human intervention, experience and expertise in some cases results in wrong diagnosis or no diagnosis. In view of this, it is considered appropriate to apply AI techniques for fault diagnosis by DGA. Use of AI for diagnosis shall enhance speed, accuracy and reliability significantly. These AI techniques are described in this article. In this article, AI techniques like Basic AI interpretation method, Artificial Neural Network (ANN) and Support vector Machine (SVM) have been discussed. 




Table – 1 (comparison of accuracies for different DGA methods using conventional techniques)

Method % of correct diagnosis % of unresolved diagnosis % of wrong diagnosis
KGM 42 0 58
RRM 62 33 5
DRM 71 26 3
DTM 96 0 4


From table – 1 it can be understood that DTM is the most efficient and accurate method.

Basic AI Interpretation Method

Recent development of AI models based on a combination of KGM, DRM, RRM and DTM techniques providing a base for more sophisticated techniques is shown in Fig – 1. This is the most basic model to understand functioning of AI for fault diagnosis. Concentration of key gases in ppm in oil samples is input to the AI logic diagram. The key gases are C2H4, CO, combination of H2 & C2H2 and H2.This AI logic diagram uses four DGA techniques i.e. KGM, RRM, DRM and DTM.

Artificial Intelligence Interpretation Method

Input is given to the KGM logic box. If the result is normal i.e. none of the key gases is present in the sample, the result is normal health of the transformer. If the result is abnormal i.e. one or more of the key gases are present in the sample, the output is fed to the DTM logic box and RRM & DRM logic box. Output of the DTM logic box shall provide information about the type of the fault.  Output of RRM and DRM logic box shall be in the form of code. If the code complies with the specified standard, output shall inform about the type of the fault. If the code is not valid as per specified standard, The output shall be ignored as unresolved.

 


Artificial Neural Network (ANN) Application To DGA

The Artificial Neural Network is a highly accurate and effective method for DGA. It is because hidden relationships between the fault types and concentration and type of dissolved gases can be recognized by ANN through training processes. The neural network technique is used to understand and categorize complex fault patterns without much knowledge about the process, the used trials or the fault patterns themselves. A neural network consists of many simple neurons which are connected with each other. An ANN is used with multilayers perceptron (MLP) and adaptive back-propagation learning algorithm for the fault diagnosis of power transformers. The architecture of the MLP is composed of multiple layers, an input layer, a variable number of hidden layers and an output layer. In particular ANN can learn and adapt to statistical data and extract essential characteristics from input data. Fault diagnosis is a process of correlating and analyzing input data records to one or more output fault conditions. ANN is a reliable, accurate and efficient tool for this analysis. The basic idea of neural network based diagnosis is non-linear mapping of inputs and outputs. For this purpose a back propagation network is generally used to diagnose the incipient transformer fault. An artificial neural network includes selection of inputs, outputs, network topology, training patterns and weighted connection of nodes. Input data pattern correctly reflects the characteristics of the fault diagnosis at output neurons. Another significant feature of the ANN design is to choose network topology. This is done experimentally through a repeated process to optimize the number of hidden layers and nodes according to the training process. We use two types of DGA methods, namely key gas method and Rogers Ratio method for fault diagnosis.  In this process two types of inputs are given. First input is 9 key gases namely hydrogen (H2), methane (CH4), ethane (C2H6), ethylene (C2H4), acetylene (C2H2), methyl ethylene (C3H6), Propane (C3H8), carbon dioxide (CO2) and carbon monoxide (CO). Second input is 3 ratios namely C2H2/ C2H4, CH4/H2 and C2H4/ C2H6. These 12 inputs are processed to diagnose 6 output parameters by ANN i.e. 5 different fault conditions, namely Partial Discharge, Arcing in oil, Arcing in solid insulation, Thermal fault in oil, Thermal fault in solid insulation and one normal (no fault) condition. In this application a Feed Forward Back Propagation network is used. A trainLM training function is used for training and adaptation of the network. MSE (Mean Square Error) is used to compute the performance measure. Block diagram for a three layer feed-forward artificial neural network is shown in Fig – 2.

Fig 2: Multi layer Neural Network

The ANN works on the TANSIG transfer function. TANSIG transfer functions calculate a layer’s output parameters processed from input parameters. 

Alternatively, output parameters may also be represented by 3 digit binary codes as shown table – 2. This shall result in reduction of output neurons from 6 to 3.

Table – 2

Output Code Type of Fault
001 Normal condition (No Fault)
010 Partial Discharge
011 Arcing in oil
100 Arcing in solid insulation
101 Thermal fault in oil
110 Thermal fault in solid insulation


ANN classification using 3 output neurons

First of all, the above mentioned faults shall be identified by the ANN network. After the test, with several parameters, the appropriate ANN architecture with optimum accuracy is obtained. Now, the optimal parameters are utilized to train the ANN model. For this purpose, following parameters are specified for an optimum design.

Three layers: an input layer, one hidden layer and an output layer

The performance of ANN network is analyzed in terms of wrong diagnosis rate and non diagnosis rate for the two DGA methods as summarized in table – 3.

Table 3


The above result indicates the superiority of the ratios method over key gas method for the ANN network classification.

The ANNs tools help the maintenance experts in planning conditioning monitoring, predictive and preventive maintenance. An important advantage of ANN based fault diagnosis is that it learns directly from the training samples and updates its knowledge when necessary. The highly non-linear mapping capability of neurons provides a comparable and often superior performance over fuzzy system solutions. ANN method is more accurately applied to Dissolved Gas Analysis since the hidden relationships between fault types and dissolved gases can be determined by ANN through a training process. 

Support Vector Machine (SVM)   Application To DGA

SVM is a machine learning algorithm that analyzes data for classification and regression analysis. SVM is a supervised learning method that looks at data and resolves it into one of the two outputs i.e. -1 or +1. The test results indicate that the SVM technique can significantly improve the accuracy for power transformer fault diagnosis.   

As shown in Figure 3, the diagnostic model comprises of six SVM classifiers which are used to identify the seven outputs i.e. normal state and the six faults as partial discharges (PD), thermal fault less than 3000C (T1), thermal fault between 3000C and 7000C (T2), thermal fault greater than 7000C (T3), low energy discharge (sparking, D1), high energy discharge (arcing, D2). With all the training samples of the output parameters, SVM1 is trained to separate the normal state from the fault state. When input of SVM1 is a sample representing the normal state, output of SVM1 is set to +1; otherwise -1. With the samples of single fault, SVM2 is trained to separate the discharge fault from the overheating fault. When the input of SVM2 is a sample representing discharge fault, the output of SVM2 is set to +1; otherwise-1. With the samples of discharge fault, SVM3 is trained to separate the high-energy discharge (D2) fault from the partial discharge (PD) and low energy discharge (D1) fault. When the input of SVM3 is a sample representing the D2 fault, the output of SVM3 is set to +1; otherwise -1. With the samples of overheating fault, SVM4 is trained to separate the high temperature overheating (T3) fault from the low and middle temperature overheating (T1 and T2) fault. When the input of SVM4 is a sample representing the T3 fault, the output of SVM5 is set to +1; otherwise -1. SVM5 is trained to separate the middle temperature overheating (T2) fault from the low temperature overheating (T1) fault. When the input of SVM5 is a sample representing the T2 fault, the output of SVM5 is set to +1; otherwise -1. SVM6 is trained to separate the partial discharge (PD) fault from the low energy discharge (D1) fault. When the input of SVM6 is a sample representing the D1 fault, the output of SVM6 is set to +1; otherwise -1.

All the six SVMs follow Polynomial and Gaussian as their kernel function. In SVM, the parameters σ and C of the SVM model are optimized by the cross validation method. The adjusted parameters with maximal classification accuracy are selected as the optimum parameters. Then, the optimal parameters are utilized to train the SVM model. So the output codification is presented in Table - 4.

Table 4
Fig – 3: SVM application for fault diagnosis


Firstly, we categories the faults by SVM with the polynomial kernel. The diagnosis results are given in Table - 5.

Table 5

Secondly, we categories the faults by SVM with the Gaussian kernel. Table – 6 lists the diagnosis results.

Table 6


According to faults diagnosis test results by the SVM with polynomial and Gaussian kernel functions, we find that the Rogers ratios method with Gaussian method is the most efficient diagnosis method. This is shown in table – 7.

Table 7

So, for comparison with other AI techniques for transformer fault diagnosis, we conclude that SVM techniques with Gaussian kernel are most efficient.


Conclusion 

In this article, the artificial intelligence techniques are implemented for faults diagnosis using the dissolved gas analysis for power transformers. The DGA methods implemented are key gas and Rogers ratios method. The fault diagnosis models performance was analyzed with artificial neural networks (ANN) and Support Vector Machine (SVM) with polynomial and Gaussian kernel functions. The real data sets are used to investigate the performance of the DGA methods in power transformer oil. The experimental results show that the SVM method with Gaussian kernel functions presents the better results in comparison to neural networks (MLP). The SVM with Gaussian kernel function provides more effective diagnostic results than the SVM with polynomial kernel function. According to test results, it is found that the ratios method is a more suitable method. The results show that the SVM with the Gaussian kernel function provides more effective results compared to any other AI methods available at present. Based on the aforesaid analysis it is appropriate to recommend the SVM method for AI based diagnosis of incipient faults and internal health of power transformers for more reliable, accurate and consistent results compared to other methods.


August 13, 2020
by 
Anupam Rastogi
Newsletter
Get great insight from our expert team.
Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
By signing up you agree to our Terms & Conditions