Leveraging Bayesian optimization and multilayer artificial neural network (MLANN) for fault prediction in oil-immersed transformers

Power transformers are crucial and high-cost equipment for the reliability and continuity of electrical systems. Consequently, regularly monitoring transformers to predict faults early based on existing parameters, is critical. The most common and well-known analysis method for fault prediction in power transformers is dissolved gas analysis, which measures the concentration of gases in the transformer oil. This research leverages a deep learning approach based on a multilayer artificial neural network with a Bayesian optimization method to enhance the accuracy of fault prediction on the dissolved gas analysis data. Following data preprocessing, which included imputation of missing values and normalization, the dissolved gas analysis dataset was divided into training, validation, and testing sets with 70 %, 15 %, and 15 % of the data, respectively. The proposed multilayer artificial neural network model, optimized using the Bayesian optimization method, achieved an outstanding accuracy of 97.99 %, pointedly outperforming benchmark machine learning classifiers. Furthermore, the model demonstrated superior performance across multiple evaluation criteria, including sensitivity, F1-score, G-mean, and Matthews Correlation Coefficient, ensuring a comprehensive assessment of predictive effectiveness. All simulations and evaluations were conducted using Python software, leveraging frameworks such as TensorFlow, Keras, Scikit-learn, Pandas, and NumPy to ensure efficient implementation. The Findings indicate that the multilayer artificial neural network classifier, which leverages the Bayesian optimization method, outperformed state-of-the-art techniques in fault prediction of oil-immersed transformers.