Development of a predictive maintenance-based anomaly prediction method to prevent unscheduled outages in power transformers

Failures occurring in power transformers within electrical distribution systems adversely affect energy continuity, resulting in reduced customer satisfaction, increased operational costs, and weakened energy supply security. This issue becomes particularly critical in regions where infrastructure is sensitive to climatic conditions. Given that a significant portion of unplanned outages is preventable, it underscores the need for identifying risks prior to failures and managing maintenance activities through a proactive approach. This thesis focuses on a novel approach aimed at predicting unplanned outages in power transformers located in the İzmir and Manisa regions by utilizing weather data, infrastructure characteristics, and historical outage records. The primary objective of this study is to anticipate potential transformer failures in advance, thereby enabling more efficient maintenance management and enhancing energy supply security. Most studies in the literature rely on either a single algorithm or limited data use, resulting in restricted scope. In contrast, this thesis comparatively evaluates multiple machine learning and deep learning algorithms and improves model performances through various optimization techniques. In the study, a comprehensive dataset was created by integrating a total of 8,671 outage records from 34 power transformers between the years 2021 and 2024, regional meteorological parameters, and transformer-specific technical attributes. Critical preprocessing steps such as addressing class imbalance, detecting outliers, and feature selection were applied; subsequently, the dataset was tested with both conventional machine learning algorithms (XGBoost, LightGBM, CatBoost) and deep learning-based models (LSTM, GRU). Model performances were evaluated using metrics such as G-Mean and ROC-AUC, and hyperparameters of the best-performing models were optimized using Optuna-based optimization methods. The obtained results revealed the type of relationship existing between weather data and failure risks and demonstrated the associations between seasonal effects and outage probabilities. This study serves as an important reference for enhancing predictive maintenance effectiveness in electrical distribution systems, reducing outage durations, and contributing to energy continuity.