Development of Multi-Layer Perceptron Model for Power Transformer Fault Identification Expert System

This study aims to develop a multi-layer perceptron (MLP) model for a power transformer fault identification expert system. The proposed system utilizes MLP to predict the presence of faults in power transformers based on Duval Pentagon Method using the dissolved gas concentration as input parameters. The system is designed to be implemented in fault diagnosis of power transformers expert system. The model is trained and tested using 1,052 and 1,051 generated data for DPM1 and DPM2, respectively. The different numbers of hidden layers, types of activation function, and solver are observed as tuning parameters. Classification accuracy, precision, recall, and F1 score are used as performance criteria. The best solver for the proposed method is lbfgs, resulting in the best hidden layer for DPM 1 being 100 with an accuracy of 0.981, and for DPM 2, using a hidden layer of 50 with an accuracy of 0.991. So, the best model is DPM 2. The proposed model is implemented in an expert system, which can help power utilities and maintenance teams in their decision-making process, resulting in more effective and efficient maintenance of power transformers.