Prediction of Distribution Transformer Faults by Logistic Regression Models Based on Cost-Sensitive Learning

Distribution transformer is the core of distribution network, and its safe operation directly affects the reliability of power supply. In this paper, the parameters such as oil temperature, oil level height and ambient temperature were first analyzed through the normality test histogram and found to be abnormal. Then, the correlation between these parameters and circuit current and voltage was explored using the Pearson correlation coefficient method, and significant correlation was found. Given that the faulty samples are much less than the normal samples, this paper introduces a cost-sensitive learning method to balance the category distribution and improve the classification accuracy without increasing the computational complexity. Finally, LDA, LR, SVM, DT and RF models are used to predict the oil temperature indicator alarm faults, and the optimal parameters are found through grid search with 5-fold cross-validation, and it is found that the logistic regression model performs the best and is more effective under the cost-sensitive learning method.