Analysis of Prediction-Based Machine Learning Application for Improving Transformer Maintenance Effectiveness

One crucial piece of electrical component whose sustainability must be taken into account is the transformer. Issues with distribution transformers in the electrical system can damage customer satisfaction and impede sales of electricity. This research aims to assess the efficacy of applying machine learning to predict the feasibility conditions of distribution transformers by using inspection data from one of the largest unit of Indonesian Electrical Company (PLN). The data from that unit contains narrower time density and evaluation from a team of field experts who have been used as a reference by management in making decisions. There are some anomalous data such as having unnatural load values, null-value phase loads and non-uniform textual data so that data cleansing is necessary. feature selection is then carried out based on the results of checking the correlation between features using the Pearson method. After obtaining the final features, some categorical data is processed using label encoding (ordinal) and one hot encoding (non-ordinal) methods to ensure that the data turns into numerical and ready to be processed into machine learning algorithms. Feature engineering is done using lagged feature technique, where the previous inspection history data is added into a new column from t-1 to t-5. From the visualization results, there is imbalanced data that needs to be followed up with the undersampling method. The Grid Search hyperparameter tuning method is used to get the best combination of parameters for the desired model and to ensure the model is not overfitting or underfitting, it is necessary to test using validation techniques, one of which uses K-Fold Cross Validation.