The Study on Power Transformer Design Optimization Using Neural Network System

Distribution transformers are most important components in a power system network, where in any distribution network a huge number of distribution transformers are used. A transformer is a device with one or two magnetically coupled electric circuits by a common magnetic field that is used to transform voltage from one level to another. Distribution transformers are also the most varied types of transformers where in the Kingdom of Saudi Arabia a huge number of transformers are installed in the public electricity supply system ranging in size from 50 kVA to 1000 kVA. Manufacturing companies have introduced computer aided transformer design that help reduce the man hour needed, but mainly to reduce and optimize the amount of material needed and to reduce the delivery cycle time to the customer. Computer design techniques will ensure that the user‘s requirements in addition to the standard design constraints such as efficiency, loaded and non-loaded losses, temperature rise and other constraints are all achieved. In the end this will give us a number of different designs which all guarantee the desired characteristics of the transformer. Even though all the designs meet the requirements, the transformers will have different parameters such as core radius, number of lamination layers, winding type and number of turns, and so on. This in turn indicates that the total cost for every transformer design will differ as the two main factors affecting the costs are; the amount of material used, and the total losses of the transformer.

TDO is a mixed-integer non-linear programming problem with a dynamic & discontinuous objective feature & constraints, with the goal of comprehensive computation of the characteristics of the transformer centered on national & foreign specifications & the specifications of the transformer customer, utilizing available materials & manufacturing techniques, to reduce manufacturing. There are some various types of accurate functions defined in the TDO literature, however the most frequently used ones are the minimization of the cost of manufacturing the transformer & minimization of the TOC that was characterized as the cost of the life- associated with the purchase and operation of the transformer. Minimization of key material costs, manufacturing costs, & TOC are the three goal feature options available in the program prepared for this report. Although the governing bodies are implementing regulations requiring the utilize of high-efficiency transformers, transforming consumers are encouraged to continue utilizing TOC since the regulations are intended solely to raise the minimum standard of transforming efficiency (Global Industry Analysts 2014.).

Artificial Neural Networks (ANN) Architecture implies how neurons are arranged in the form of layers and the type of interactions between the neurons. Architectures of neural networks are closely related to the training/learning algorithm applied on these networks. There are three broad categories of neural network architectures. (a) Single Layer Feed forward ANN (b) Multilayer Feed forward ANN (c) Recurrent Networks Single Layer Feed forward ANN have a single layer in the whole network. It consists of input layer which is called the zeroth layer and the other layer is the output layer which is the single layer. The signals are passed from input layer to output layer and not in the reverse direction and there are no closed paths. Therefore these networks are called as Feed forward networks. Figure 1(a) depicts the Single Layer Feed forward Artificial Neural Networks. Multilayer Feed forward Artificial Neural Networks have more than one layer. The layers between the input & output layers are known as the hidden layers. Due to these hidden layers, the artificial neural networks are also known as Multi-Layer Artificial Neural Networks. The signals from the input layer are transmitted to hidden layer which are further passed to another hidden layer till the last layer known as the output layer. The signals are not passed in the reverse direction from output layer towards the input. Due to this property of a network, these networks are termed as Multi-Layer Feed forward Artificial Neural Networks. Figures 1 (b to e) show various types of Multilayer Feed forward Artificial Neural Networks. The Multilayer Feed forward ANN which use sigmoid ac tivation functions at the hidden layer nodes are termed as Sigmoidal Feed forward Artificial Neural Networks. Recurrent Neural Networks (RNN) have feedback loops, that is, the computation flow is bi-directional. For example, Hopfield Networks have feedback loops from every node to every node, and is a completely connected network. Elman and Jordan Networks (which are used for temporal analysis) have feedback loops from higher layers to lower layers, while competitive networks have lateral connections. See Figure 2 for RNN architectures. Feedback loops means that outputs of the network are fed back to the inputs in a single layer network and in networks which also have hidden layers feedbacks can be from outputs of the hidden layers to previous hidden layers or/and input nodes or the feedbacks can be from outputs of the network to hidden layers and/or input nodes. variants depending upon the problems these networks solve like function approximation, forecasting and control, pattern association and optimization.

Power Transformer is the nerve center of any distribution system. Generally speaking, the ability of power transformers is determined from the charge survey data. The electricity transformers are the substation's primary machinery, while other machinery is concerned with the transformers' practical aspects. Any transformer fault will have Buchholz relay and differential relay operated. The transformer will not be paid with working on such relays. It must be isolated, & tests should be carried out. The unit should only be put into service if the healthiness is confirmed by tests. The potential faults in power transformers are usually the following: • Within the transformer tank faults • On heat treatment • Outside transformer region errors (due to fault) • Initial faults • Internal faults (For fixing these faults Differential safety schemes shall be provided) • Switch to fault, Arcs inside the Oil(Sudden stress relay schemes are given to resolve such kinds of faults) For addition, transformer faults can be separated into two major groups for review purposes. They are: I by faults, i.e. by overloads and by external short circuits. (ii) External defects, i.e., transformer winding & contact failures. External faults may be sub-divided into starting faults & severe faults.

While BP can provide a very concise distributed analysis of various data sets, it does have its drawbacks such as slow learning, broad training sets, simple sticking to local minima and weak robustness. A variety of approaches have been suggested to strengthen the BP norm in this way. Genetic algorithms, wavelet transforms and neural networks based on particle swarm intelligence are becoming promising alternatives amongst them. This has been a topic of active research in the recent past and the open literature contains many good methods. Since ANNs can provide outstanding pattern recognition, several researchers recommend them for power transformer relaying implementation. Neural network research must be treated with special caution. The multilayer perceptron with controlled back-propagation learning algorithms is often utilized in developing new security strategies. For power transformer security systems, however, a hybrid wavelet transforms & neural network (WNN) approach and PSO-trained ANN method is being used in the research work.

An ANFIS in which each node executes an incoming signal unique operation (node function) & series of parameters relating to that node. Conceptually, there are virtually no limitations on an adaptive network's node functions beyond differentiability piecewise. Structurally, the only disadvantage in network setup is that the form in feed forward will be this. Because of these minimal restrictions, the applications of the adaptive network in various fields are immediate & immense. The current framework is referred to as ANFIS, which stands for Fuzzy Inference Scheme based on the Adaptive-Network. Throughout this case, a hybrid learning rule is suggested that combines the gradient method & estimation of the least squares (LSE) to define parameters. The variable set method of

Wavelet transforms is an useful tool for simultaneously extracting information from the current signals in the time & frequency domains. This is simply a flexible windowing strategy. Such flexible window duration requires long time periods to be utilized when desiring more precise low frequency information & quick time intervals while desiring high frequency information. Figure shows a time series signal for Wavelet transformation. Analysis Wavelet adopts the concept of scale & connections between scale & frequency. Discrete Wavelet Transform (DWT) is commonly used with the analysis of non-static signals.

Discrete wavelet transformations (DWT) allow for the use of dyadic scales and positions, inferring scales and positions based on a two-probable power. The DWT will establish fine-scale coefficients to capture data of high recurrence and the coarse-scale coefficient to capture data of low recurrence. Typically, the DWT technique is used to capture the signal distress. The intent behind prominence was its ability to set an immense magnitude of the coefficients to zero without substantial data malfortune. In addition, if more properties are required than fixed properties of a signal, DWT will also most likely be a good alternative to conventional Fourier transform (FT) strategy.

Eleftherios I et al. (2006) Transformer architecture is basically calculated by reducing general production costs, including commodity costs and labor costs. In every event, this minimization has to rely on the limitations set by external determinations and the wishes of clients. In this article, an groundbreaking theory similar to the Decision Tree approach is proposed that power transformers can be constructed using just ten simple input parameters. The method is modified by programming. The resulting kit is appropriate for consumers who are not experts in the field of transformers and, however, for transformer manufacturers who want a robust and favorable

accordance with various approaches that are impossible to succeed in a potential solution in the initial run. In fact, the production meetings of the transformer are integrated into this particular system, which allows only the novice to create the perfect configuration of the transformer. The suggested method and programming consists of a simple unit that is currently widely used in the manufacturing of transformers. ELEFTHERIOS I. AMOIRALIS et al. (2008) In this article, a robust enhancement approach is introduced with the goal of reducing the complex component expense of wound center allocation transformers, taking into consideration the limitations faced by both external data and company requirements. In order to be carried out in these lines, Mixed Integer Programming relevant to Branch and Bound Strategy is used. There is a comparison between the current approach and the heuristic development program of the industrial transformer sector, and the findings demonstrate the power and predominance of this modern methodology. Ali Soldoozy et. al. (2018) Numerous improvement equations and methods have been suggested during the last decades, most of which are in some form or another influenced by design and by regular circumstances, in particular those that can be essentially investigated and formulated. The goal of this work is to propose another streamlining strategy, inspired by the growth and fertility of plants, that enables them to prune a few sections of the plant or tree, i.e., horrible sections of the pursuit space that do not meet the constraints of the issue and quickly scan the whole research space. Through partnership, we will prune the shooting area of the improvement problem, cut and forget its awful pieces, pieces that definitely do not follow the requirements of the question, with the goal that the speed and consistency of the inquiry will be significantly enhanced. Through enhancing the standard of shooting, we hope to fully dispose of getting shot at the very minimum. There are subtle comparisons with pruning a tree to increase its diversity and pruning a shooting field for an development problem to boost the efficiency of the measurement of enhancement. JMAG-Designer technology is used for the study of finite components. N Aishwarya et al. (2019) Established social orders currently require an ever-increasing supply of electrical power, and the requirement has risen last year. Complex systems are developed to satisfy this increasing demand. The efficient functioning of such a complex system typically relies on the ability of that structure to include a strong and continuous inventory of loads. Throughout the key requirement but, throughout addition, it is the responsibility of the power system engineers to provide the buyers with steady and reliable control. Such problems require the need to examine the soundness of the control system. Throughout this article, a consistent state security investigation is carried out using swing conditions and the knowledge gathered from the investigative methodology is used to train the Artificial Neural Network(ANN) such that a reliable state stability status structure is established. Sudhaet al. (2019) Scenario preparation is one of the most important activities faced by large-scale strategy and operation engineers. Power grid engineers use a contingency model to analyze the design of the network and to determine the need for additional transmission extensions due to the change in load or age. The various approaches used to evaluate such contingencies rely on the complete AC load flow analysis or the reduced load flow or responsiveness variables. In any scenario, such techniques need an immense amount of computing resources and are not appropriate for on-line applications in broad power frameworks. It is difficult to make up-to-date on-line contingency research using traditional methods in the context of the disagreement between the quicker method and the consistency of the arrangement. Accordingly, this paper suggests a computationally efficient technique for contingency research utilizing a synthetic neural system. Mahesh Yanagimath et. al. (2019) Power transformers are the central components of electrical power transfer and control systems. They are continually under the control of physical, hydraulic, warm and normal anxieties. It is the most important and expensive parts of the control system. Owing to the expense of the power transformer, the monitoring and maintenance of the state of the transformer is becoming important. There are various predictive approaches necessary for the transition of the well-being situation. One technique named ANN is used to evaluate and differentiate blames for oil-filled power transformers that are based on Dissolved Gas in Oil Analysis (DGA). By using the ANN feed forward, the most severe precision can be obtained with the aid of the back spread algorithm. This paper presents the problem description structure for the ANN power transformer.

Throughout this research, Genetic Algorithms (GA), Particle Swarm Optimization (PSO) & Teaching Learning Based Optimization (TLBO) have been utilized to solve the Transformer Design Optimization (TDO) problem. The No-Free Lunch may be assumed that a certain meta-heuristic might have very good results on a certain set of problems, yet the same algorithm may have bad results on a specific set of problems. Thus, in this segment, the success of three separate artificial intelligence strategies (i.e. GA, PSO and TLBO) will be evaluated to solve TDO problems.

This section explains the methods utilizing GA for optimum design of 100 kVA, 11/0.433 kV, propagation transformer. The key benefit of GA is that a minimal change in the software multiple target functions may be streamlined. The set of decision variables chosen for 100 kVA, 11/0.433 kV, Dyn-11 delivery transformer as seen in Table 1.1 after consultation with a local transformer supplier.

This section explores the findings obtained by GA, PSO and TLBO in the optimisation of transformer architecture. Table 1.1 describes the restrictions on judgment factors, whereas the constraints are just like those mentioned in Table 1.1. For statistical inferences, each optimization routine was run 20 times to track the optimum global out of many local minima. Optimization Results with GA Every decision variable in this study is coded as a 20 bit binary string in GA. Optimum size measurements & efficiency parameters are extracted from the analysis as objective function of a collection of GA parameters & penalty factor with active component rate. To find the global optimum, 3 distinct RWSBGA techniques, SRWSBGA & TSBGA were used.

In this article, each optimization algorithm was run for 500 generations. However, no improvement in objective function was observed when the numbers of iterations were increased beyond 100. Table 1.4 and Table 1.5 indicate the value of optimal design variables of a 100 kVA, 11/0.433 kV distribution transformer obtained by RWSBGA, SRWSBGA, TSBGA, PSO and TLBO after performing 20 trial runs. Results indicate that PSO and TLBO were able to find a better value of the objective function as compared to GA . After conducting 20 trial runs with different initial random values, TLBO was

standard deviation.

The final design dimensions & performance parameters of 1-star & 2- star rated transformers found out by PSO & TLBO and their comparison with the Exhaustive Search Method is depicted in Table 1.6, while the dynamic performance characteristics of transformers designed by TLBO is shown in Table 1.7.

Transformers carry on a crucial job of connecting power systems at different voltage levels. It would actually not be possible, without the transformer; to use electric power from different perspectives is used today. Genetic algorithms were used to develop transformers and minimize work costs, just as they were used to prepare the neural network for power misfortune prediction in transformers. The transformer structure optimization problem depends on minimizing or amplifying a target function that is exposed to a couple of imperatives. Between various target functions, the target functions normally used include minimizing absolute mass, minimizing complex component costs, minimizing specific material costs, minimizing assembly costs, minimizing all out reporting costs or boosting transformer rated power.

Ali Soldoozy & Abdolreza Esmaeli (2018) ―Implementation of tree pruning method for power transformer design optimization‖, RESEARCH ARTICLE, Int Trans Electr Energ Syst. 2019;29:e2659. https://doi.org/10.1002/etep.2659 wileyonlinelibrary.com/journal/etep Amit Kumar Yadav, Abdul Azeem, Hasmat Malik (2011). ―Optimization of Power Transformer Design using Simulated Annealing Technique‖, International Journal of Electrical Engineering. ISSN 0974-2158 Volume 4, Number 2 pp. 191-198 © International Research Publication House. ELEFTHERIOS I. AMOIRALIS*, PAVLOS S. GEORGILAKIS, MARINA A.( 2008) ―Design optimization of distribution transformers based on mixed integer programming methodology‖, JOURNAL OF OPTOELECTRONICS AND ADVANCED Eleftherios I. Amoiralis, Pavlos S. (2006). ―Methodology for the Optimum Design of Power Transformers Using Minimum Number of Input Parameters‖, ICEM, PAPER NUMBER 470 Jafferson J. M. and Hariharan P. (2013). Machining performance of cryogenically treated electrodes in microelectric discharge machining: a comparative experimental study. Mater. Manuf. Process. 28(4): pp. 397–402 Mahesh Yanagimath (2019). ―Diagnosis of Power Transformer by using Artificial Neural Network‖, International Journal of Scientific & Engineering Research Volume 10, Issue 5, May- ISSN 2229-5518 N Aishwarya1,Manjula S (2019). ―Analysis of Steady State Stability of Power System using Artificial Neural Network‖, International Journal of Scientific & Engineering Research Volume 10, Issue 5, May- ISSN 2229-5518 J. Saied, N. Reza, S. Mahdi and B. Saeid (2011). ―PSS and SVC controller Design using chaos, PSO and SFL algorithm to enhancing the power system stability‖, journal of energy and power engineering, Vol. 3, pp. 87-95. K. M Sangram and R. Kuwaresh (2012). ―Improving power system stability using real-coded Genetic Algorithm based PI controller for STATCOM‖, international journal of power system operation and energy management, Vol. 1, pp. 73-78. P.H. Sasongko, F.M. Rain and I.W Hatta (2013). ―Multi-Objective function GA for modal optimal control design in PSS and UPFC power oscillation Damping coordination‖, Journal of theoretical and applied information technology, Vol. 51, pp. 472-481. J. Seyyed, R. Reza and K. Hamed: ―Application of Genetic algorithm to design PID controller for Power system stabilization‖, Center of Excellence for mechatronics, University of Tabriz Tabriz, Iran. Sudha, Manjula S. Sureban (2019). ―Contingency Analysis of Power System Using Artificial Neural Networks‖, International Journal of Scientific & Engineering Research Volume 10, Issue 5, May-2019 ISSN 2229-5518

Research Scholar, Sunrise University, Alwar, Rajasthan