A Model for Predicting Transformer Failure Tendencies on 132 kilovolts Power Network in Uganda

Abstract

Uganda’s 132 kV transmission network has experienced increasingly frequent and costly power transformer failures, often occurring without warning and leading to cascaded outages, extended downtime, and significant corrective maintenance expenses. The prevailing asset management framework remains largely reactive, lacking an effective condition-based monitoring and predictive maintenance strategy. This study was thus motivated by the need to develop a robust prediction tool capable of identifying transformers approaching critical degradation, minimizing unexpected failures, and guiding timely replacement decisions based on actual transformer health rather than age alone. To achieve this, the study set out to: (i) identify key operational parameters influencing 132 kV transformer failure, (ii) develop a predictive model for degradation under diverse stochastic and loading environments, (iii) validate the model’s accuracy, and (iv) derive a cost-effective replacement strategy. Using a quantitative research approach, historical condition-based monitoring data for 30 transformers across selected substations was collected over a 23-year period. A hybrid Particle Swarm Optimization– Adaptive Neuro-Fuzzy Inference System (PSO-ANFIS) model was developed to predict transformer degradation trends using key condition variables such as breakdown voltage (BDV), moisture content, acidity, interfacial tension (IFT), and dissolved gas analysis (DGA) parameters. Model performance was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The PSO-ANFIS model consistently outperformed the traditional ANFIS across all parameters, with average improvements of 25–35% in accuracy. For BDV, the PSOANFIS achieved an RMSE of 1.75, MAE of 1.42, and MAPE of 5.4%, compared to ANFIS values of 2.35, 1.89, and 7.2%, respectively. Similar improvements were observed for other indicators—moisture (MAPE 4.7% vs. 6.8%), acidity (5.2% vs. 8.5%), and key DGA gases (average MAPE 5.5% vs. 8.4%). The results showed that the PSO-ANFIS model significantly improved prediction accuracy, enabling early identification of asymptomatic transformers and supporting a proactive, cost-efficient replacement strategy tailored to the Ugandan grid. The study’s findings present an important step toward modernizing transformer asset management using intelligent prediction models. xii