You are on page 1of 10


Presented By
N.Bhuvaneswari III/IV E.E.E K.Lakshmi III/IV E.E.E

A.M.Reddy memorial college OF Engineering& technology

Petlurivaripalem GUNTUR Dist.

ABSTRACT This paper presents as Artificial Neural Network (ANN) technique to recognize the incipient faults of power transformers. The technique presented in this paper improves the diagnosis accuracy of the conventional dissolved gas analysis (DGA) approaches. The ANN is trained by using Adaptive Back-propagation learning algorithm that converges much faster than the conventional Back-propagation algorithm. The developed ANN system for the power transformer fault diagnosis has superior performance in fault diagnosis has superior performance in fault diagnosis as compared to the conventional methods. INTRODUCTION The main function of a power system is to supply electrical energy to its customers with an acceptable degree of reliability and quality. Among many other things, the reliability of a power system depends on trouble-free functioning of power transformers. Consequently, their maintenance and particularly their preventive maintenance can lead to huge savings, besides achieving uninterrupted power supply. There have been routing maintenance schedule available for electrical engineers for the maintenance of power transformers but there has been no reliable diagnostic test to assess the internal condition of the transformers except the dissolved gas analysis (DGA) test. The DGA of transformer oils is a widely accepted technique used as a diagnostic tool for detecting the incipient faults in power transformers. Early detection of fault results not only in large savings in operation and maintenance costs but also prevents any premature breakdown/failure besides improving the overall system reliability. The cost DGA of an oil sample is about Rs 500 only whereas the cost of even a 20/25 MVA power transformer is above fifty lacs and, if it fails, its repairs cost would be 10% to 20% of its cost. By the diagnosis and rectification of incipient faults, this huge amount can be saved. In the event of local overheating , partial discharge or arc discharge in a high voltage transformer, its insulting materials will be decomposed into the gases. The decomposed gases include hydrocarbons (machine, ethane, ethylene and acetylene) and other gases (hydrogen, carbon dioxide etc). The decomposed gases will be dissolved into the transformer oil. These gases are then

extracted from the oil under high vacuum and analyzed by Gas Chromatograph to get each as concentration separately. By interpretation of gas contents, the developing faults in the power transformers can be diagnosed. Many diagnostic criteria1-3 have been developed for the interpretation of these gases. By these methods, one can find relationships between the gases and the fault conditions, some of which are obvious and some of which may not be apparent (hidden relationships). However, much of diagnostic relies on experts to interpret the results correctly. The key gas method and the ratio methods are commonly used for the interpretation of dissolved gases. The Artificial Neural Network method can be used more accurately for this purpose since the hidden relationships between the fault types and dissolved gases can be recognized by ANN through training process. An ANN with Adaptive Back-propagation learning algorithm for the fault diagnosis of power transformer is presented in this paper. DESIGN OF ANN An Artificial Neural Network included selection of inputs, outputs, network topology and weighted connection of node. Input feature-selection constitutes an essential first step. This is chosen very carefully so that the input features will correctly reflects the characteristics of the problem. Another major task 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 training performance and prediction accuracy. In the current study, five key gases H2, CH4, C2H6, C2H4, and C2H2 are chosen as input features. Since overheating, partial discharge and arcing are the three major fault types in power transformers; hence there will be four output patterns to be identified including the normal condition.


For many years there was no theoretically sound algorithm for training multi-layer ANN, and therefore, the applications of ANN were severely limited. The invention of Back propagation algorithm has played a vital role in the resurgence of interest in ANN. Backpropagation is a systematic method for training multi-layer ANN. It has a strong mathematical foundation. A two-layer feed forward network is shown in Figure 1. It consists of number of neurons connected by links divided into two layers. A set of inputs is applied from outside or from previous layer. Each of these is multiplied by a corresponding weight w. The sum of the weighted inputs and the bias b forms the input n to the transfer/activation function F. Neurons may use any differentiable, monotonic increasing transfer functions to generate their outputs. Back-propagation networks often use the log-sigmoid
and tan-sigmoid transfer functions. For each neuron in the first layer, the neuron output is given by:

And output of the second layer is:

Where p is the input to the first layer; F1 and F2 are the transfer function of first and second layers respectively, b1 and b2 are the biases of first and second layer respectively and w1 and w2 are the connection weights of the first and second layers respectively. The network is trained to learn the relationships between the inputs and target outputs. For training, a number of pairs of input patterns p and target patterns t are presented to the ANN and then the ANN is asked to adjust weights in all connecting links and also the biases in the nodes such that the desired output patterns are produced at output nodes. In general, the network output a2 will not be same as the target or desired values, i.e., t. For each pattern, the Sum Square of the Error is:

The main goal of back-propagation (BP) algorithm is to adjust the connection weights w and biases b to minimize error between desired output and actual network output. A common approach to achieve this goal is the generalized delta rule (GDR). To achieve convergence towards improved values for the weights and biases, the incremental changes in w and b can be calculated by using following equations:

Where h is the learning rate constant.


Exponential smoothing. The method involves adding a term to the weight adjustment that is proportional to the amount of previous weight change.

Back-propagation network is very slow because it requires small learning rate for stable learning. Sejnowski and Rosenberg (1987) described a method for improving the training time of back-propagation algorithm based on where s the smoothing coefficient in the i range of 0.0 to 1.0. s the learning rate constant. Training time is further decreased by the i use of Adaptive learning rate, which attempts to keep the learning step size as large as possible while keeping learning process stable. In the current study, the ANN is trained using the Adaptive back-propagation learning algorithm as described above. This algorithm consists of repeatedly passing the training sets through the neural network until its weights and biases minimize the output error the entire set of inputs. The learning rate () is updated during training process. First, the initial network output and error are calculated. At each epoch, new weights and biases are calculated using the current learning rate. New output and error are then calculated. If the new exceeds the old error by more than a predetermined ratio (1.05), the new weights, biases, output and error are discarded. In addition, the is decreases (by multiplying it by 0.74), otherwise the new weights etc are kept. If the error is less than the old error, the is increased (by multiplying it by 1.05). The complete training procedure is explained in flowchart as shown in Figure 2. DIAGNOSIS RESULTS A two-layer feedforward ANN with 60 neurons in the hidden layer is designed and trained by using 30 training patterns obtained from different power transformers having capacity from 6 MVA to 250 MVA of Rajasthan Rajya Vidhyut Prasaran Nigam (previously known as Rajasthan State Electricity Board). The desired error goal is set to 0.001, and is obtained in 208 training epochs as compared to 5000 epochs to get error goal of 0.012 in BP algorithm. After successful

Training, the weight matrices and biases are stored as files. The testing patterns are input from a data file and then the ANN is used to detect the faults. Ten testing samples from different power transformers are tested. The testing patterns and corresponding results are listed in Table 1 and Table 2, respectively. A comparative statement of fault diagnosis of transformes by various methods in listed in Table 3. This table clearly shows that the results of actual reports (as received from Oil Testing Laboratory/Repair shop at 400 KV G S S, Jaipur) and the ANN diagnosis match for all patterns except pattern number 7.

CONCLUSION In this paper, a new transformer-fault-diagnosis-system has been proposed and implemented successfully to improve the conventional DGA methods. Theoretically, the ANN can be trained to represent any observable phenomenon if there are sufficient data available. The more complex the relationship is, the more training data are needed. The use of Adaptive Back-propagation learning algorithm reduces drastically the number of training epochs required for desired accuracy as compared to conventional Backpropagation algorithm. The proposed approach also enables new and inexperienced engineers to diagnose easily whereas the conventional diagnostic methods can only be used by experts. REFERENCES 1. F W Heinrichs. The Impact of Fault-Detection Methods and Analysis on the Transformer Operating Decision. IEEE Transactions on Power Delivery, vol 2, PWRD-2, no 3, July 1987, p 836. 2. O Vanegas, Y Mizuno, K Naito and T Kamiya. Diagnosis of Oil-insulated Power Apparatus by Using Neural Network Simulation. IEEE Transactions on Dielectrics Electrical Insulation, vol 4, no 3, June 1997, p 290. 3. Y Zhang, X Ding, Y Liu and P J Griffin. An Artificial Neural Network Approach to Transformer Fault Diagnosis. IEEE Transactions on Power Delivery, vol 11, no 4, October 1996, p 1836. 4. P D Wasserman. Neural Computing Theory and Practice. Van Nostrand Reinhold, New York, 1989. 5. B Yegnanarayana. Artificial Neural Networks. Prentice-Hall of India Private Limited, 1999.