Fault Diagnosis of Power Transformers Based on Comprehensive Machine Learning of Dissolved Gas Analysis

Currently, the commonly used fault diagnosis methods of power transformers are often difficult to deal with the ambiguity problems encountered in the troubleshooting process. Even with some artificial intelligent techniques already experimented, the results have not been not systematic compared and are far from real application. Therefore, this paper tries to make a thorough use of machine learning tools towards result data from dissolved gas analysis. This paper establishes a machine learning model based on dissolved gas analysis for internal fault diagnosis of power transformers, and makes a comparison between multiple machine learning methods. Firstly, an overall neural network model of transformer diagnosis based on the dissolved gas analysis is formed. Next, the performance of optimized BP neural network, probabilistic neural network (PNN), and decision tree algorithm is compared from the aspects of speed and accuracy. Furthermore, the case-based reasoning method based on the Euclidean distance and normalized energy intensity algorithm is employed to get the closest matching similar case to realize the transformer defect prediction and assistant decision-making. With actual examples verified, the case-based reasoning method can help detect the most likely abnormal causes of faulty transformers through providing the most similar matching case.