Transformer No-Load Losses Calculation Using Machine Learning

This paper explores the innovative application of machine learning to calculate no-load losses in transformers. Existing calculation methods are either insufficiently accurate, as they do not consider all factors contributing to additional losses, or computationally slow due to the complexity of finite element method models. The accurate determination of these additional losses is crucial for a correct transformer design. The proposed calculation method offers a fast and highly accurate prediction of losses, utilizing training data consisting of measured values of manufactured transformers, providing the optimal source for modeling reality. Gaussian process regression is chosen as the learning technique because of the limited number of samples. Three surrogate models are presented, each providing a single output value that represents the percentage of additional losses to the nominal core loss. For further validation, the results of the surrogate models are compared with the analytical calculation. The nominal core loss can be calculated as the product of the core weight and the specific loss. The first model incorporates four input variables: nominal core loss, leg pinch, window height, and core cross-section area. The second model expands to include three additional inputs: steel grade, calculated flux density, and maximum steel sheet width. Both models employ variables with continuous kernels. Because the second model with seven inputs predicts new data slightly less accurately, a third similar model is introduced. In this third model, one variable, steel grade, utilizes a categorical kernel due to its specific behavior. This minor adjustment improves the accuracy of the prediction compared to the test data. Moreover, all presented models exhibit significantly higher accuracy compared to the analytical calculation.