| 1. IntroductionPower transformer is the most important and most expensive transmitting and transforming electricity equipment, it leads to most of the accidents in electric power system. Along with the growing need of electricity in the country and the trend that to be a experimental unit in northeast area, reliable power supply and prolonged equipment longevity are desiderated in the electric power system. Accurate and timely fault diagnosis methods are demanded imminently. Oil and paper are usually used to insulate and emit heats in large power transformers. Under the condition of normal running and early faults, the transformer oil and the paper insulated material will decompose to produce a few low molecular hydrocarbons such as firedamp (CH4), ethane (C2H6), ethene (C2H4), ethane (C2H2), carbon monoxide (CO), carbon dioxide (CO2) and hydrogen (H2). Most of the gases will dissolve in the oil. In a certain extent, the component and content of dissolved gases in transformer indicate the insulation aging and faults degree of transformer. Dissolved Gases Analysis (DGA) is one of the main technology methods to diagnose the internal faults in transformers. This method uses the component, content and producing speed of the dissolved gases in transformer oil. In long-term research and practice of transformer fault diagnosis, scientists have summarized methods using the gases content or the gases content ratio. In these years methods of fuzzy mathematics and expert system using DGA were carried out, but these methods have some disadvantages and localizations.Due to the producing complexity of the dissolved gases in transformer oil, transformer faults are the results of transformers themselves and their synthesized actions of application environment and long-term accumulation. The transformerfault symptoms are various, and the relationship between fault symptoms and fault mechanics are also complicated, which make it hard to find out the connection between the component and content of the transformer oil dissolved gases and the transformer fault types and conditions using traditional theories. So it is very difficult to establish a transformer fault diagnosis method. Artificial Neural Network has many advantages, such as parallel distributed processing, self-adapted ability, association, memory, clustering, faults tolerance etc. It supplys a new way of acquiring knowledge, expression and illation, it can show out some mapped relation through finding out the optimal weights by training different information seriatim, which made it a proper method for the multi-process, multi-fault and multi-mode transformer fault diagnosis.2. Research ContentBased on the research and analysis of present and normal transformer fault diagnosis methods, study on the power transformer fault diagnosis using BP (Back Propagation) Neural Network theory is done in the paper. Main research contents are as follows:(1) Research of the Relationship between Fault Types and Dissolved Gas.The capability of the transformer oil and solid insulation as well as the producing principle of the dissolved gases are analyzed systemically in the paper. Transformer inner representative faults types along with the corresponding gas components and characteristics are introduced. Deep study of the methods that can diagnose transformer fault characters and types have been made. Moreover, common diagnosis methods have also been introduced, such as character gas method, three-ratio method, non-coding ratio method and the assistant methods for integrative diagnosis. 150 groups of typical and representative dissolved gases in transformer oil analysis data have been collected, and 36 groups have been chosen by testing.(2) Deep Study of BP Neural Network Model, Algorithm and Common Improved Algorithms.Through the analysis of BP Neural Network model, algorithm and network learning rules, several BP algorithm common improved algorithms, such as variable learning rate backpropagation, momentum backpropagation, levenberg-marquardt and resilient backpropagation are deeply studied. Three improved algorithms for increasing network constringence and advancing network capability are present in the paper, which are variable learning rate backpropagation with network error-based self-regulating amending genes, variable learning rate combines momentum backpropagation with network error-based self-regulating amending genes and resilient backpropagation with double factors-based self-regulating amending genes.(3) The Transformer Fault Diagnosis Model Based on BP Neural Network. According to the transformer fault gases and the fault types, a type of 5-12-5 BP Neural Network model for transformer fault diagnosis is established on the basis of designing network structure, pretreating data and determining transfer function.(4) Improved Algorithms for Increasing Network Constringence and Advancing Network Capability.①A new improved algorithm—variable learning rate backpropagation with network error-based self-regulating amending genes is present. The principles of the algorithm are as follow:During the neural network training process, the learning rate amending genes are self-regulated by the changing process of neural network errors to increase amending extent, decrease useless iteration and make the learning rate adapted to the network convergence. By this way, the largest accepted learning rate can be used in learning process, which can improve the network convergence speed.The established neural network is trained by this improved algorithm. To compare with other algorithms, the network is also trained by several other common algorithms. Network training results have showed that compared with the original algorithm, the training epochs are decreased by 35.4% and the convergence speed is increased by 44.9%, the relative coefficients are above 0.995 between network outputs and expectable outputs, the network square sum errors is 0.000999567 while the goal is set to 0.001. The results explain that this algorithm can effectively improve the neural network convergence speed and decrease the training epochs. This algorithm has great theoretical research significance and far-ranging utilizing value.②Another new improved algorithm—variable learning rate combines momentum backpropagation with network error-based self-regulating amending genes is present. The principles of the algorithm are as follow:This algorithm utilizes not only the variable learning rate backpropagation with network error-based self-regulating amending genes to make the learning rate amending genes self-regulated by the changing process of neural network errors, decrease useless iteration and increase amending extent, but also the momentum backpropagation to increase the extent of weights and bias. This algorithm can effectively improve the neural network convergence speed and ensure steady training process.The established neural network is trained by this improved algorithm. To compare with other algorithms, the neural network is also trained by several other common algorithms. Network training results have showed that compared with the original algorithm, the training epochs are decreased by 6.6% and the convergence speed is increased by 18.6%, the relative coefficients are above 0.994 between network outputs and expectable outputs, the network square sum errors is 0.000992859 while the goal is set to 0.001. The results explain that this algorithmcan improve the network convergence speed and decrease the training epochs a certain extent.③The third new improved algorithm—resilient backpropagation with double factors-based self-regulating amending genes is present. The principles of the algorithm are as follow:Double factors in this algorithm are the gradient direction of two continuous iteration and the network total square errors difference of two continuous iteration. During the neural network training process, the weights and bias amending genes are self-regulated by the changing process of neural network errors and gradient to increase amending extent. The weights and bias amending extent will be decreased when concussion appears. The weights and bias amending extent will be increased when weights and bias are amended on the same gradient direction. The algorithm will eliminate influence of the gradient extent to improve the neural network convergence speed, retain steady training process and decrease the possibility of concussion.The established neural network is trained by this improved algorithm. To compare with other algorithms, the network is also trained by several other common algorithms. Network training results have showed that compared with the original algorithm, the training epochs are decreased by 45.1% and the convergence speed is increased by 36.4%, the relative coefficients are above 0.996 between network outputs and expectable outputs, the network square sum errors is 0.00096617 while the goal is set to 0.001. The results explain that this algorithm can effectively improve the network convergence speed and decrease the training epochs. This algorithm has great theoretical research significance and far-ranging utilizing value.(5) Transformer Fault Diagnosis Based on Several Improved BP Algorithms. After training, the aquired neural network model is used for transformer faultdiagnosis. The fault diagnosis accuracies are 90%, 90% and 95%, which accord with the actual condition of the power transformer fault diagnosis.3. ConclusionBased on the deep research of present and normal transformer fault diagnosis methods, a BP Neural Network model for transformer fault diagnosis is established, three new improved algorithms of BP algorithm are present and used for training network and fault diagnosis based on the founded model. The conclusions in the paper are as follows:(1) According to the transformer fault gases and the fault types, a type of 5-12-5 BP Neural Network model for transformer fault diagnosis is established. After training and fault diagnosing, the model's fault diagnosis accuracy is above 90%, which shows that the model is proper, feasible and correct for the transformer fault diagnosis.(2) Through simulation and fault diagnosis, it has showed that the three algorithms present in this paper are correct and ascendant, they can effectively improve the network convergence speed and the model's capability. The acquired transformer fault diagnosis model has comparatively high accuracy and good application effect, which accords with the actual condition of the power transformer fault diagnosis. The algorithms present in the paper can supply reference basis for the further research of transformer fault diagnosis based on the neural network and have a comparatively high generalization for other neural network research fields. |
PDF Full Text Request