Transformer No-Load Losses Estimation Using Machine Learning From Measured Data

This work introduces an innovative application of established machine learning methods to calculate transformer no-load losses. An accurate estimation of losses is crucial to cost-effective and reliable transformer design. Existing methods often lack precision due to hard-to-predict additional losses or are computationally expensive due to complex finite element models. The proposed method enables fast and accurate prediction of transformer no-load losses using Gaussian process regression, selected for its suitability for the dataset and problem. Building on previous work, an improved surrogate model is developed with enhanced outlier filtration and optimized input selection. The approach is further extended to predict a range of losses, accounting for manufacturing deviations. Its performance is validated through a learning curve. Key findings indicate that outlier removal significantly improves accuracy and the optimal model depends on nine key variables from an extended input set. Modeling ranges of losses is feasible with a sufficiently large dataset, which further enhances accuracy, as confirmed by the learning curve. Overall, the method shows a strong potential for accurate no-load loss estimation and to support more efficient and cost-effective transformer design.