Abstract:- Power transformer is one of the most important and expensive equipment in electrical network. The transformer oil is a very important component of power transformers. It has twin functions of cooling as well as insulation. The oil properties like viscosity, specific gravity, flash point, oxidation stability, total acid number, breakdown voltage, dissipation factor, volume resistivity and dielectric constant suffer a change with respect to time. Hence it is necessary that the oil condition be monitored regularly to predict, if possible, the remaining lifetime of the transformer oil, from time to time. Six properties such as moisture content, resistivity, tan delta, interfacial tension and flash point have been considered. The data for the six properties with respect to age, in days, has been taken from literature, whereby samples of ten working power transformers of 16 to 20 MVA installed at different substations in Punjab, India have been considered. This paper aims at developing ANN and ANFIS models for predicting the age of in-service transformer oil samples. Both the the models use the six properties as inputs and age as target. ANN (Artificial Neural Network) model uses a multi-layer feedforward network employing back propagation algorithm, and ANFIS (Adaptive Neuro Fuzzy Inference System) model is based on Sugeno model. The two models have been simulated for estimating the age of unknown transformer oil samples taken from generator transformers of Anpara Thermal Power Project in state of U.P. India. A comparative analysis of the two models has been made whereby ANFIS model has been found to yield better results than ANN model.
Keywords: ANN, ANFIS, Power Transformer, Regression, Performance, Backpropagation Algorithm
Power transformer is one of the most important constituent of electrical power system. The transformer oil, a very important ingredient of power transformers, acts as a heat transfer fluid and also serves the purpose of electrical insulation. Its insulating property is subjected to the degradation because of the ageing, high temperature, electrical stress and other chemical reactions. Hence it is necessary that the oil condition be monitored regularly. This will help to predict, if possible, the in-service period or remaining lifetime of the transformer oil, from time to time.
There are several characteristics which can be measured to assess the present condition of the oil. The main oil characteristics are broadly classified as physical, chemical and electrical characteristics; some of these are viscosity, specific gravity, flash point, oxidation stability, total acid number, breakdown voltage, dissipation factor, volume resistivity and dielectric constant. There exists a co-relation among some of the oil properties and suffer a change in their values with respect to time [2]. This variation of oil properties with respect to time has been utilised to develop the two models as said earlier
The training data for the proposed work have been obtained from literature, whereby ten working transforms of 16 to 20 MVA, 66/11 KV installed at different substations in the state of Punjab, India have been considered. The six properties of transformer oil such as breakdown voltage (BDV), moisture, resistivity, tan delta, interfacial tension and flash point have been considered as inputs and age as target. Test data have been taken from generator transformers of 250 MVA, 15.75kV/400kV from Anpara Thermal Power Project in state of U. P., India.
II. “Ann” and “Anfis” methods
It is known that classical models need linear data for their processing, therefore models like ANN and ANFIS that are based on soft computing techniques, play an important role for solving these kinds of non-linear problems.
Neural networks exhibit characteristics such as mapping capabilities or pattern association, generalization, robustness, fault tolerance, parallel and high speed processing. Neural networks can be trained with known examples of a problem to acquire knowledge about it. Once trained successfully, the network can be put to effective use in solving unknown or untrained instances of the problem. ANN model which uses multilayer feed forward network is based on back propagation (BP) learning algorithm of neural network. Backpropagation gives very good answers when presented with inputs never seen before. This property of generalization makes it possible to train a network on giving set of input-target pairs and get good output.
ANFIS stands for Adaptive Neural Fuzzy Inference System. Using a given
input/output data set, the toolbox function ANFIS constructs a fuzzy inference
system (FIS) whose membership function parameters are tuned (adjusted) using
either a backpropagation algorithm alone, or in combination with a least
squares type of method. This allows the fuzzy systems to learn from the data
they are modelling. These techniques provide a method for the fuzzy modeling
procedure to learn information about a data set, in order to compute the
membership function parameters that best allow the associated fuzzy inference
system to track the given input/output data. This learning method works
similarly to that of neural networks.
III.
Development of ann model
The proposed ANN
model uses “Levenburg-Marquardt
(trainlm) algorithm which is independent of learning rate, hence by simply changing
the number of neurons in hidden layer, training and testing error could be
reduced. A total of 700 data sets obtained from
literature [2] were arranged in tabular form and used for training the neural
network. The model uses a simple
two layer network, one hidden layer and one output layer. Input layer comprises
of six neurons, one for the each input, while the output layer has a single
neuron for a single output, the age of oil sample.
It has been found that network architecture
that uses 20 neurons in hidden layer gave the best performance with a
regression of 0.999 and mean square error (MSE) of 83.0 ( data is non
–normalized, so error looks large ) . The training continued for 184 iterations
with training functions logsig in hidden layer and purelin in output layer
respectively.
