Multi frequency ultrasonic detection of water content in transformer oil with GA-BPNN

摘要

Oil is an important liquid insulating medium in oil-immersed transformer. The water content in oil is closely related to the deterioration of the performance of the insulation system. 160 oil samples that were collected from various transformers, 150 of which were set as training sets and 10 forecast sets. Input variables to the model is 242-dimension multi-frequency ultrasonic testing data, and the output of the model is the water content of the oil sample. Through the experimental method to determine the hidden layer consists of 11 neurons, the nonlinear mapping relationship. Genetic algorithm (GA) was applied to optimize the BP neural network connection weights and threshold of every layer. The results show that the correlation coefficient of water content of oil prediction model is 0.97207. The average absolute percentage error MAPE of the proposed GA-BPNN model is 9.4%. All results provide a new online detection method for water content in transformer oil.