Proceedings of the 8th WSEAS International Conference on ELECTRIC POWER SYSTEMS, HIGH VOLTAGES, ELECTRIC MACHINES (POWER '08) Transformer oil’s service life identification using neural networks L. EKONOMOU1 1 2 P.D. SKAFIDAS2 D.S. OIKONOMOU3 Hellenic American University, 12 Kaplanon Street, 106 80 Athens National Technical University of Athens, 9 Iroon Politechniou St., 157 80 Athens 3 Agricultural University of Athens, 75 Iera Odos Street, 118 55 Athens GREECE Abstract: In this paper a neural network method is developed and presented for high and medium voltage transformer oil’s service life identification. Mineral oils are used as insulating and cooling agents in transformers due to their good aging behaviour and low viscosity. Their in-service characteristics are continuous varied with time having as a result the degradation of the insulation and the decrease of the residual operating time of transformer oil. In this paper an alternative to up today method is presented in order to identify transformer oil’s service life based on neural networks (NN). In contrast to the traditional time consuming and expensive measurements a NN model is presented, trained with actual recorded data, capable to identify with accuracy the transformer oil voltage breakdown and consequently its service life. The NN model is applied on operating high and medium voltage transformers presented very accurate results. The proposed approach can be useful in the work of electrical maintenance engineers reducing maintenance time and cost. Key-Words: Dielectric characteristics; multilayer perceptron; neural networks; oil voltage breakdown; service life; transformer oil. on-line and off-line monitoring-diagnosis systems [4]. 1. Introduction Power transformers are integral components of power systems. One of the factors that can reduce the efficiency of power systems is their failure, creating significant cost implications, unscheduled supply interruptions and in very extreme cases even explosions. These undesirable consequences are mainly due to the degradation of the insulation of the transformer oil, which is an essential means for the electrical insulation and heat dissipation of transformers [1]. Therefore is obvious that the accurate identification of transformer oil’s residual operating time can help in the more effective management of transformer’s maintenance schedule and in the increase of system’s reliability. The current work presents a new approach for the transformer’s oil service life identification based on artificial intelligence and more specifically neural networks (NN). NN present to give solutions with significant results in almost every power systems problem varying from load forecasting [5] and power systems transient stability [6], to prediction of the magnetic performance of strip-wound magnetic cores [7] and high voltage transmission line problems [8], exploiting NN computational speed, ability to handle complex non-linear functions, robustness and great efficiency even in cases where full information for the studied problem is absent. The safe and prolonged operation of power transformers is generally provided through time consuming and expensive field measurements related mainly to the transformer oil’s breakdown voltage, acidity and water content. Alternative methods have been proposed based on polynomial regression models [1], measurements on oil’s dielectric properties (complex permittivity and tanδ) [2], gas chromatography [3] and integrated ISSN: 1790-5117 A NN model capable to identify with accuracy the transmission oil’s breakdown voltage - and consequently its service life - is developed based on actual field data and is applied on operating high and medium voltage transformers presented very accurate results. The proposed approach can be useful in the work of electrical maintenance 222 ISBN: 978-960-474-026-0 Proceedings of the 8th WSEAS International Conference on ELECTRIC POWER SYSTEMS, HIGH VOLTAGES, ELECTRIC MACHINES (POWER '08) In order to train the network, a suitable number of representative examples of the relevant phenomenon must be selected so that the network can learn the fundamental characteristics of the problem. More than a hundred different learning algorithms are available, but the most popular one is the backpropagation. In a backpropagation NN the learning algorithm has two phases. First a training input pattern is presented to the network input layer. The network then propagates the input pattern from layer to layer until the output pattern is generated by the output layer. If this pattern is different from the desired output, an error is calculated and then propagated backwards through the network from the output layer to the input layer. The weights are modified as the error is propagated [9, 10]. engineers reducing the maintenance time and cost, offering greater efficiency in power systems. 2. Neural networks (NN) A neural network consists of a number of very simple and highly interconnected processors, called neurons, which are analogous to the biological neurons in the brain. The neurons are connected by weighted links that pass signals from one neuron to another. Each link has a numerical weight associated with it. Weights are the basic means of long-term memory in NN. They express the strength, or importance, of each neuron input. A neural network “learns” through repeated adjustments of these weights. The characteristic feature of these networks are that they consider the accumulated knowledge acquired during training and respond to new events in the most appropriate manner, giving the experience gained during the training process. In this work a typical neural network model known as multilayer perceptron model (MLP) has been used (Fig. 1). The MLP is a feed forward NN with an input layer of source neurons, at least one middle or hidden layer of computational neurons, and an output layer of computational neurons. The input layer accepts input signals from the outside world and redistributes these signals to all neurons in the hidden layer. The hidden layer detects the feature. The weights of the neurons in the hidden layers represent the features in the input patterns. The output layer establishes the output pattern of the entire network [9]. NN models are determined according to their architecture, i.e., the network’s structure, transfer function and learning algorithm. In the backpropagation networks the MLP architecture is generally decided by trying a) varied combinations of number of hidden layers and number of nodes in a hidden layer, b) different transfer functions and c) learning algorithms in order to be selected the most suitable NN model architecture, which has the best generalizing ability amongst the all tried combinations [10, 11], i.e., minimization of the sum-squared error. Once the training process is completed and the weights and bias of each neuron in the neural network is set, the next step is to check the results of training by seeing how the network performs in situations encountered in training and in others not previously encountered. 3. Transformer oil Mineral oils are fabricated by refining a fraction of the hydrocarbons collected during the distillation of petroleum crude stock. Their good aging behaviour and low viscosity make them good insulating and cooling agents. Furthermore they are wide available and have a relative low cost. These characteristics have as a result the extensive use of them as transformer oils [1-4]. However there are instances of transformers failing whilst in service, creating significant cost implications for power suppliers and in extreme cases explosion with a consequent threat of workers for severe injury or death and significant environmental impacts. Therefore continuous monitoring of residual operating time of Fig.1 MLP feed forward neural network. ISSN: 1790-5117 223 ISBN: 978-960-474-026-0 Proceedings of the 8th WSEAS International Conference on ELECTRIC POWER SYSTEMS, HIGH VOLTAGES, ELECTRIC MACHINES (POWER '08) transformer oil is required something that is conducted worldwide from all electric companies based on real measurements. The most popular measurements for this purpose are the transformer oil breakdown voltage, the acidity and the water content measurements, which either solely or in combination with the others can conclude on the degradation of the oil’s insulation [1, 3]. As it has mentioned earlier each NN model is determined according to its structure, the transfer function and the learning algorithm, which are used in an effort to learn the network the fundamental characteristics of the examined problem. The learning algorithms and the transfer functions are used to adjust the network’s weights and biases in order to minimize the sum-squared error. The structure of the networks i.e. the number of hidden layers and the number of nodes in each hidden layer, is generally decided by trying varied combinations for selecting the structure with the best generalizing ability amongst the tried combinations. In general one hidden layer is adequate to distinguish input data that are linearly separable, whereas extra layers can accomplish nonlinear separations [10]. This approach was followed, since the selection of an optimal number of hidden layers and nodes is still an open issue. 4. Design of the proposed NN The goal is to develop a neural network architecture that could identify the oil breakdown voltage of high and medium voltage transformers. Identifying the oil breakdown voltage conclusions on the oil’s service life can be deduced. Five parameters that play important role to the transformer’s oil breakdown voltage were selected as the inputs to the neural network, while as an output the oil voltage breakdown value was considered. These data constitute actual collected/measured data from measurements performed by Hellenic Public Power Corporation S.A. [12] and National Technical University of Athens [13]. In this work several different multilayer perceptron models were designed and tested in order to be identified the model with the best generalizing ability. These were combinations of five different learning algorithms, five different transfer functions and several different structures consisted of 1 to 3 hidden layers with 2 to 50 neurons in each hidden layer (table 2). Table 1 presents the proposed NN model architecture for the assessment of transformer’s oil voltage breakdown. The five used input parameters are: the transformer type (medium or high voltage), the transformer service period (in months), the total acidity (in mg KOH/g oil), the relative density value of the oil and the water content (in ppm). The output parameter of the NN model is the transformer’s oil voltage breakdown. It must be mentioned that the authors have selected only those five input parameters for the developed NN model, although there are also others that affect the transformer’s oil voltage breakdown. The reason was that there were not available records of data for any other parameters. Table 2. Designed MLP NN Models Structure - 1 to 3 hidden layers Learning Algorithm - Gradient Descent - 2 to 50 neurons in each hidden layer - LevenbergMarquardt Transfer Function - Hyperbolic Tangent Sigmoid - Logarithmic Sigmoid - Quasi-Newton - Hard-Limit - Conjugate - Competitive Gradient - Random Order - Linear Incremental Table 1. ANN Architectures Input Variables - transformer type T Output Variables - oil voltage breakdown B 5. NN Training, Validation and Testing - service period S - total acidity A The MATLAB neural network toolbox [14] was used to train the neural network models. Two hundred and one values of each input and output data, referring to two hundred and one different - relative density D - water content W ISSN: 1790-5117 224 ISBN: 978-960-474-026-0 Proceedings of the 8th WSEAS International Conference on ELECTRIC POWER SYSTEMS, HIGH VOLTAGES, ELECTRIC MACHINES (POWER '08) be mentioned that transformer oil breakdown voltages for those ten samples has been also determined according to the IEC testing method [13] in order to compare them with the NN assessed values. Fig. 2 presents the comparison of the actual/measured transformer oil’s breakdown values and these obtained using the designed NN model. It is obvious that the results obtained according to the proposed NN method are very close to the actual ones, something which clearly implies that the proposed NN model is well working and has an acceptable accuracy. Fig. 3 presents the percentage error between actual measured transformer oil’s breakdown voltage and NN estimation values for each one of the ten Hellenic operating transformers. The average percentage error for these 10 transformers is 8.5 %. measurements in 150 kV and 400 kV transformers, were used to train and validate the neural network models. In each training iteration 20% of random samples were removed from the training set and validation error was calculated for these data. The training process was repeated until a root mean square error between the actual output and the desired output reaches the goal of 1% or a maximum number of epochs (the presentation of the set of training data to the network and the calculation of new weights and biases), it was set to 10,000, is accomplished. Finally, the transformer oil’s breakdown voltage value was checked with the number obtained from situations encountered in the training and others which have not been encountered. Following the training, validation and testing process of all possible combinations of structure, transfer function and learning algorithm, it was selected and used further to assess the transformer oil’s breakdown voltage the model, that presented the best generalizing ability, had a compact structure, a fast training process and consumed low memory. This NN model had the following characteristics: two hidden layers, with sixteen, and thirteen neurons in each hidden layer respectively, gradient descent learning algorithm and hyperbolic tangent sigmoid transfer function. 20,00% Precentage error between actual measurements and NN results 15,00% 10 10,00% 5,00% 0,00% 1 2 3 4 5 6 7 8 9 10 -5,00% -10,00% 9 -15,00% Transformer Samples 8 Transform er sam ples 7 Fig. 3: Percentage error between actual measured transformer oil’s breakdown voltage and NN estimation values for ten Hellenic operating transformers. 6 5 4 3 2 1 6. Conclusions 0 5 10 15 20 25 30 35 40 45 Transformer oil's breakdown voltage (kV) Actual measurements NN estimation Fig. 2: Comparison of the actual transformer oil’s breakdown voltage and the NN estimation values for ten Hellenic operating transformers. The selected NN model has been used in order to identify the transformer oil’s breakdown of ten different operating high and medium voltage transformers of the Hellenic power system. It must ISSN: 1790-5117 225 The paper describes a neural network method for transformer oil’s identification of high and medium voltage transformers. Actual input and output data measured from over 200 Hellenic operating transformers were used in the training, validation and testing process. The developed NN model is applied on ten different operating transformers of known transformer oil’s breakdown voltage, presenting great accuracy. The proposed NN approach can be useful in the studies of electrical maintenance engineers since from the transformer’s oil breakdown voltage easily can be deduced safe conclusions on oil’s service life. ISBN: 978-960-474-026-0 Proceedings of the 8th WSEAS International Conference on ELECTRIC POWER SYSTEMS, HIGH VOLTAGES, ELECTRIC MACHINES (POWER '08) References: [1] M.A.A. Wahab, M.M. Hamada, A.G. Zeitoun, G. Ismail, Novel modeling for the prediction of aged transformer oil characteristics, Electric Power Systems Research, Vol. 51, 1999, pp. 61-70. [2] C.T. Dervos, P. Vassiliou, P. Skafidas, C. Paraskevas, Service life estimation of transformer oil, Int. Conf. on Protection and Restoration of the Enviroment VI, Skiathos, Greece, 2002, pp. 1239-1246. [3] V.G. Arakelian, Effective diagnostics for oilfilled equipment, IEEE Electrical Insulation Magazine, Vol. 18, 2002, pp. 26-38. [4] D. Nedelcut, D. Sacerdotianu, G.Tanasescu, S.Nicolae, L. Voinescu. On-line and off-line monitoring-diagnosis system (MDS) for power transformers, International Conference on Condition Monitoring and Diagnosis, Beijing, China, 2008, pp. 949-955. [5] B. Kermanshahi, H. Iwamiya H, Up to the year 2020 load forecasting using neural nets, Electric Power Energy Systems, Vol. 17, 2002, pp. 789-797. [6] A.D.P. Lotufo, M.L.M. Lopes, C.R. Minussi, Sensitivity analysis by neural networks applied to power systems transient stability, Electric Power Systems Research, Vol. 77, No. 7, 2007, pp. 730-738. [7] G.K. Miti, A.J. Moses, Neural network-based software tool for predicting magnetic performance of strip-wound magnetic cores at medium to high frequency, IEE Proc-Sci. Meas. Technol., Vol. 151, No. 3, 2004, pp. 181-187. [8] L. Ekonomou, V.T. Kontargyri, St. Kourtesi, T.I. Maris, I.A. Stathopulos, Artificial neural networks in high voltage transmission line problems, Institute of Physics Publishing, Measurement, Science and Technology, Institute of Physics Publishing, Measurement, Science and Technology, Vol. 18, No.7, 2007, pp. 2239-2244. [9] [11] R. Lippmann, An introduction to computing with neural nets. IEEE ASSP Magazine, Vol. 4, No. 2, 1987, pp. 4-22. [12] Public Power http://www.dei.gr . S.A., [13] C.T. Dervos, C.D. Paraskevas, P. Skafidas, P. Vassiliou, Dielectric characterization of power tranformer oils as a diagnostic life prediction method, IEEE Electrical Insulation Magazine, Vol. 21, 2005, pp. 11-19. [14] H. Demuth, M. Beale, Neural Network Toolbox: For use with MATLAB, The Math Works, 1994. Lambros Ekonomou was born on January 9, 1976 in Athens, Greece. He received a Bachelor of Engineering (Hons) in Electrical Engineering and Electronics in 1997 and a Master of Science in Advanced Control in 1998 from University of Manchester Institute of Science and Technology (U.M.I.S.T.) in United Kingdom. In 2006 he graduated with a Ph.D. in High Voltage Engineering from the National Technical University of Athens (N.T.U.A.) in Greece. His research interests concern high voltage engineering, transmission and distribution lines, lightning performance and protection, stability analysis, control design, electrical drives and neural networks. Panagiotis D. Skafidas was born in Kalamata, Greece in 1964. He received his B.Sc. degree in electrical engineering from the National Technical University of Athens (NTUA), 1989. He also obtained the Ph.D. degree in electrical engineering from NTUA, 1994. He is currently as a part time teaching assistant lecture in the Department of Information Transmission Systems and Material Technology of NTUA. His research activities cover surface physics, high voltage insulators, dielectrics and neural networks. Dimitrios S. Oikonomou was born on September 28, 1982 in Athens, Greece. He received his degree in Natural Resources Management and Agricultural Engineering from the Agricultural University of Athens in Greece. He is currently a research assistant in the Agricultural University of Athens. His research interests concern neural networks, modelling, simulation, electrical motors and electrical conversion. M. Negnevitsky, Artificial intelligence: a guide to intelligent systems, 2nd Edition, Addison-Wesley, 2005. [10] K. Hornik, Some new results on neural network approximation. Neural Networks, Vol. 6, 1993, pp. 1069-1072. ISSN: 1790-5117 Corporation 226 ISBN: 978-960-474-026-0