Study of Over-Sampling Methods Used in Distribution Transformer Connectivity Verification

Data-driven method has been used to carry out distribution transformer connectivity verification. The imbalanced dataset always results in bad performance of machine learning algorithms. In order to solve this problem, two kinds of over-sampling methods have been studied in this paper. Voltage curves of 3967 distribution transformer which belong to 197 10kV feeders have been collected. The performance of over-sampling methods under different sampling rate has also been studied. Pareto optimality has been carried out in order to obtain the best sampling rate. Results show that with the increase of sampling rate, the value of true negative (TN) increase and the value of true positive (TP) and accuracy decrease. The performance of synthetic minority over-sampling technique (SMOTE) is a little better compared with simple copy method (SCM). When the lower limits of TN, TP and accuracy are set as 0.93, the best sampling rate of SCM is 17 and the values for SMOTE are 16, 17 and 18.