Study On Breakdown Voltage Of Large Generator Ground Wall Insulation Based On BP Neural Networks

Large generators are the main equipments in large power plant and the hearts in the power system. The growing demand of electrical energy in modem society and the rapid development of power industry have resulted in a higher requisition for the safe generating operation of generator units. Stator winding in a large generator is exposed to a combination of thermal, electrical, vibrational and thermo-mechanical stresses in service. In the long term, these multiple stresses cause insulation aging which leads to final filature. In the same time, if generator wall insulation is damaged, it will cost much time to repair or replace, so it will cause a bad effect. Therefore, studying on residual lifetime aimed at large generator ground wall insulation has important practical meanings for safe generating operation of power system.By far, residual breakdown voltage is considered to be the most objective criterion for remaining life or insulation condition. Since breakdown test is destructive, it is very expensive and unwise to make experiment like that. So it is strongly expected to predict the residual breakdown voltage by employing nondestructive diagnostic parameters. There are five kinds of nondestructive diagnostic parameters for large generator ground wall insulation: dc characteristics, dc characteristics, dielectric characteristics, partial discharge characteristics and non-electrical characteristics. After analyzing, the paper has drawn a conclusion thatdielectric characteristics and partial discharge characteristics are fit for predicting the residual breakdown voltages. So this paper analyzed the correlation between the nondestructive diagnostic parameters and the residual breakdown voltages by using the ready data with Pearson(x, y), in the last, four parameters were gained.The relation between the nondestructive diagnostic parameters and the residual breakdown voltages is very complex; the relation can't be described by an accurate mathematics model. To solve this problem, this paper has used the BP neural network that is proved to be good at dealing with the nonlinear problem. A typical BP neural network is composed of "input layer", "hided layer" and "output layer". The neural cells are connected by each other in the different layers, but disconnected in the same layers. To predict the residual breakdown voltages, the networks should be trained. To train the networks, the nondestructive diagnostic parameters and the residual breakdown voltages should be input into the BP neural network and networks parameters should be chosen, when the train begins, the neural cells are activated, the activated values are propagated from the "input layer" to the "output layer" through the "hided layers", then, output responses are gained in the neural cells in the "output layer". Then, the output responses return to the "input layer" from "output layer" to the "input layer" through the "hided layer" in the direction in which the error between the expected residual breakdown voltages and the real residual breakdown voltages can decrease. With this process recycling, in the last, the expected residual breakdown voltages will be equal to the real residual breakdown voltages, then the networks come to balance and stop the train. Now, this trained neural network can be used to predict the residual breakdown voltage with the nondestructive diagnostic parameters. In this article, the maximum imitated relative error is 1.99%, and the minimum imitated relative error is -0.65% through the chosen neural network; So from above, it can be drawn a conclusion that the imitation was very successful. So it is feasible of BP neural network to predict the residual breakdown voltages by using the nondestructive diagnostic parameters.This paper has studied a method of predicting the residual life of large generator ground wall insulation, this method has two step mainly: firstly, Thenondestructive diagnostic parameters that have more close relation with the residual breakdown voltages are be chosen ; secondly, the BP networks is be trained with the chosen nondestructive diagnostic parameters. This method has practical meaning for advancing the work of valuing and predicting residual breakdown voltage of large generator ground wall insulation, decreasing the cost and workload, saving work time, realizing the on-line monitoring for valuing and predicting residual breakdown voltage of large generator ground wall insulation.