Research on Machine Learning Method for Ultrasonic Temperature Measurement of Transformer

When the discharge or current loss occurs at the insulation defect location of the oil-immersed power transformer, it will inevitably generate heat and form a local hot spot. Too high hot spot temperature will accelerate the insulation aging process, and even lead to the loss of the original insulation performance of the transformer. Therefore, nondestructive and accurate diagnosis of transformer winding hot spot fault location is of great significance. On the basis of summarizing the research results of transformer hot spot temperature measurement at home and abroad, this paper put forward a detection method of transformer hot spot location based on ultrasonic, and used COMSOL to build a twodimensional model of transformer and ultrasonic sensor. The received time, sound pressure value and other data of each sensor were obtained through simulation, and five characteristic parameters were extracted from each received ultrasonic signal, so as to generate the sample database of the algorithm model. Further, the intelligent learning algorithm was used to fit the nonlinear mapping relationship between the hot spot location information and the characteristic parameters to form the location network of transformer winding hot spot fault. The results show that the positioning error of BP neural network model was 6.67%, and its diagnosis accuracy was significantly better than that of support vector machine (SVM) and decision tree. The research is based on machine learning algorithm to accurately identify the location of winding hot spot fault, and realize the online nondestructive monitoring of oil-immersed power transformer, which is of great significance to ensure the safety, stability and economy of power system operation.