| 1. IntroductionThe power transformer is not only the key electrical equipments, but also one of the most accident-prone equipments in the electric system. Its operating state has a tremendous impact on the safety of the power systems. In addition, it is a significant issue for electrical department to find potential faults of the transformer so as to keep it operating safely. Therefore, the fault diagnosis technology is available and reliable to operate and maintain the transformer. Transformer oil is usually used for insulation and emitting heat in large power transformers nowadays. Transformer oil and solid organic insulative material will deteriorate gradually under the work voltage then they will be decomposed to produce a few low molecular hydrocarbons such as firedamp (CH4), ethane (C2H6), ethene (C2H4), ethine (C2H2) and carbon monoxide (CO), carbon dioxide(CO2), hydrogen(H2) because of the electricity, heat, oxidation and part electric arc etc. These gases are almost dissolved in the transformer oil. The inner over- heat-fault or discharge fault will quicken the speed of gases producing. The component and capacity of dissolved gases can reflect the degree of insulation aging or transformer fault in certain extent, so it can be used as characteristics which can reflect power equipment abnormity. 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. Characteristic gases analysis, Three-rations-method etc. are the methods commend by GB/T7252-2001《Guide to the analysis and the diagnosis of gases dissolved in transformer oil》. With the development of artificial intelligence technology these years, fuzzy mathematics, grey theory and expert systems etc. based on DGA are used in transformer condition evaluation and fault diagnosis, but all these methods are have their own problems.Artificial Neural Network has many advantages, such as parallel distributed processing, self-adapted ability, association, memory, clustering, faults tolerance etc. and it is a proper method for the multi-process, multi-fault and multi-mode transformer fault diagnosis. But ordinary BP algorithm's learning is a gradient descent method, so it has inevitable shortcomings such as slow convergence and entrapment in local optimum etc. Genetic algorithms (GA) can implement global search in complex, multi-peak, non-linear and non-realization space. And it dose not require gradient information of error function. Genetic Algorithm is good at global search and BP neural network is more effective in the local search. So the algorithm, which combines genetic algorithm and error back-propagation neural network (BPNN) is applied to diagnose the fault of transformer, is a viable method.2. Research ContentBased on the research and analysis of present and normal transformer fault diagnosis methods, this dissertation establish a hybrid algorithm which combines niche genetic algorithms with BP neural network base on DGA and use it to diagnosis the faults of power transformers. 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 is analyzed systemically in the dissertation. Transformer inner representative faults types along with the corresponding gas components and characteristics are introduced. Deeply studies 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. 180 groups of typical and representative dissolved gases in transformer oil analysis data have been collected.(2) Research of the principles of BP neural networks. This dissertation studies on artificial neuron model, as well as artificial neural network topology and network learning rules. And analyze the defects of BP neural network. Several BP algorithm common improved algorithms such as momentum back-propagation and variable learning rate back-propagation are deeply studied.(3)Research of the principles of Genetic Algorithms and Niche Genetic Algorithms. This dissertation studied the principles of Standard Genetic Algorithms and the process of manipulation. And detailedly introduces the genetic operations about selection, crossover and mutation. The principle and steps about Niche Genetic Algorithms (NGA) were deeply studied. And this paper compared the performance of Niche Genetic Algorithms with Standard Genetic Algorithms using Shubert test function. The result shows that the Niche Genetic Algorithms is valid.(4) According to the previous text, this paper presents a improved algorithm model based on NGA-BP which makes good use of searching virtue in overall range using genetic algorithm and great capability of searching in local range using error back-propagation algorithm to against the defects of BP neural network. The model has two parts. One part is BP neural network and another is Niche Genetic Algorithms.The part of BP neural network: According to the transformer fault gases and the fault categories, a type of 5-10-5 BP Neural Network model is established on the basis of designing network structure, preprocessing data and determining transfer function.The part of NGA: according to the structure of BP neural network, the gene vectors are encoded with connection weights and thresholds of the neural networks, and ascertained the fitness function. Then, to search the global most optimization individual by using the operator such as roulette wheel selection and optimization saved and adaptive crossover and adaptive mutation and niche pass etc. When we find the most optimize individual, we decode it to connection weights and thresholds of the BP neural networks.(5)Train the NGA-BP model. The NGA-BP model which was established in the dissertation optimizes the weight and thresholds of the designed neural networks through the niche genetic algorithm. And then find the optimization initial connection weights and thresholds. It can improve the speed and stability of convergence of network. The simulation results of NGA-BP network model show that the speed of convergence and stability of convergence have greatly increased compared with BP network model. The rate of NGA-BP network convergence is increased by 17.59% compared with BP network when adopting steepest descend back-propagation. The rate of NGA-BP network convergence is increased by 27.65% compared with BP network when adopting momentum back-propagation. And the rate is increased by 24.74% when adopting variable learning rate back-propagation. At the aspect of Convergence stability, the successful rates of NGA-BP network are all over 90% when using the three different training methods. There are 91% while using steepest descend back-propagation and 93% while using momentum back-propagation and 98% while using variable learning rate back-propagation. Corresponding rates of BP network are 74% and 81% and 93% respectively.(6) Transformer Fault Diagnosis Based on NGA-BP model. After training, the NGA-BP model neural network is used for transformer fault diagnosis with the data which are collected. The fault diagnosis accuracies is 90%, which accord with the actual condition of the power transformer fault diagnosis.3. ConclusionBased on the deeply analysis of the traditional power transformer fault diagnosis methods and research of artificial neural network and genetic algorithms which are used for power transformer fault diagnosis, a algorithms which Using genetic algorithm to optimize the network connecting value and threshold value of BP neural network is present in this paper. And a NGA-BP model for power transformer fault diagnosis is established. Then train the NGA-BP network and fault diagnosis base on the founded model. The conclusions in the paper are as follows:1. The NGA-BP model which is established in this paper used for transformer fault diagnosis. The model's fault diagnosis accuracy is 90%, which is satisfied the requirements of fault diagnosis of power transformer. The simulation results show that the model is proper, feasible and correct for the transformer fault diagnosis.2. In this paper, the proposal method improved not only the speed of network convergence but also the stability of network convergence very well. The proposal method also improve the efficiency of fault and the precision of fault detection. The algorithms presented in the paper can supply reference basis for the further research of transformer fault diagnosis based on genetic algorithms and neural network and have a comparatively high generalization for other neural network research fields. |
PDF Full Text Request