Last modified: 2023-07-10
Abstract
Electricity is one of the essential energy sources in our daily lives and industrial activities. The electricity generated at the power plants is transmitted at an ultra-high or a high voltage, gradually reduced to a lower voltage through power transformers, and ultimately delivered to customers. In the electric power distribution system, transformers play an important role in transmission and transformation facilities. Early detection of internal faults in a transformer is crucial for preventing sudden power outages and ensuring a stable energy supply.
Dissolved gas analysis (DGA) is one of the most widely used diagnostic tools for detecting and evaluating faults in oil-insert transformers [1] [2]. Currently, there are various machine-learning technologies used for detecting faults in transformers. However, most of the datasets that can be collected are usually imbalanced. That is to say, there are a lot of data in the normal state, but there is relatively little data in the fault state, and the dataset is imbalanced.
Until now we have confirmed the identification accuracy of AI technology exceeds that of conventional DGA methods [3]. In this research, we applied multi-layer perceptron (MLP) and random forest (RF) for classification. To solve the issue of an imbalanced dataset, we applied several countermeasures such as over-sampling, under-sampling, synthetic minority technique (SMOTE), and focal loss function. Moreover, we compared the prediction accuracy between the single-step and multistep approaches. The results showed that the multistep approach achieved high prediction accuracy.