Research On BP - NN Based On Improved GA Optimization In Fault Diagnosis Of Power Network

With the rapid development of economic, The development of China’s power industry towards the direction of greater and higher level of automation. Grid stability requirements become higher. If the power grid malfunction, it will bring huge losses to the power companies and the users. It is an important means to ensure the grid operation safely and reliably is diagnose the power grid failure effectively. The first step is diagnose the faulty components of grid and isolation the faulty components, and take measures to restore the power supply, so that recovery the work of grid quickly and reduce downtime. However, it is still an unresolved problem to diagnose grid failure quickly and accurately, and it is more difficult when the protection and circuit breaker malfunctions or multiple failures. In this thesis mainly studies area diagnosis of the protection components of grid, by studying Genetic Algorithm and Back Propagation Neural Network, when regional component unusually, find the fault area and improve the power grid fault diagnosis fault tolerance.It is an effective means to analyze grid failure by diagnosing grid components. Due to the neural network has parallel distributed processing, adaptive, associative memory and clustering and so on, it apply to the whole grid diagnose components and regional diagnose components. Currently, BP neural network is one of the most extensive neural network model applied to the fault diagnosis of the grid. BP neural network has good self-learning, adaptive ability and generalization ability, but BP neural network algorithm is based on gradient method, easy to fall into local minimum value, network convergence will be slow and convergence accuracy is not high, or even non-convergence when the number of learning samples are numerous, and complex relationship between input and output. GA has the capability of global optimization, it can effectively improve BP neural network convergence speed and accuracy, while taking advantage of genetic algorithm to optimize the initial weights threshold of BP neural network and avoid BP neural network fall into local minimum, improve the effectiveness of fault diagnosis. However, traditional genetic algorithm is not perfect, in the process of global optimization, it is easy to get local optimal solution, We called "premature". In order to improve global optimization ability of the genetic algorithm, and overcome the "premature" issue, we introduced chain competitive strategy based on the traditional genetic algorithm, produce population diversity by feature selection, so that easy to achieve global optimization results.The main research work of this paper:divided into six chapters, The first chapter, it describes the purpose and significance of research, research before grid fault diagnosis, the application of GA-BP network algorithm in fault diagnosis of grid; The second chapter describes the principles and characteristics of genetic algorithms and the improved Genetic Algorithm, because the traditional genetic algorithm is easy to fall into the "premature", so that it get local optima, and it is difficult to escape from local to global optimum, we can use some strategies to strengthen the global search ability, so that avoid falling into local optimum; The third chapter introduces the principles and characteristics of the BP neural network and the selection method of hidden layer nodes in BP neural network, it is importance to select the hidden layer nodes, there must be an optimum number of hidden nodes. Then, according to some summary formula of previous research references to BP neural network, narrow the range of hidden layer nodes, then set up a neural network, the integer in the range as the number of hidden nodes, and training samples, record every training result and comparing these results, find the best hidden layer nodes in the range; The fourth chapter describes combination of improved genetic algorithm and the BP neural network algorithm containing determine the optimum number of nodes in the hidden layer; the fifth chapter, the improved algorithm of this paper is applied to troubleshooting of grid area, and simulation comparison with the traditional GA-BP network algorithm and a single BP neural network algorithm; The sixth chapter is a summary and outlook section.