A paper by authors from R&S Group and Rauscher & Stoecklin, published for CIGRE Session 2026, examines how physical modelling and machine learning can support more accurate prediction of no-load losses in distribution transformers.
The study compares two modelling approaches using measurements from more than 12,000 transformers, highlighting both the predictive precision of machine learning and the continuing value of physics-based methods.
Modelling No-Load Losses in Distribution Transformers: A Comparison Between Physical and Machine Learning Approaches
Authors: Mehrdad Kiani-Oshtorjani, Gökhan Kalkan, Axel Kirchner, François Dehlas and Rolf Fluri.
Publication: CIGRE Session 2026, Paper A2-12607, Study Committee A2: Power transformers and reactors.
Abstract
Accurately estimating no-load losses is a key challenge in transformer design, as these losses are influenced by complex nonlinear magnetic phenomena and their secondary effects. Although numerical simulations can provide high accuracy predictions, they are computationally expensive and time-consuming, which limits their use in iterative optimization workflows. In this study, a refined physics-based loss-separation model and a data-driven model based on extreme gradient boosting algorithm are developed and validated using a dataset of more than twelve thousand measured transformers operating at 50 Hz. The physical model achieved a mean absolute error (MAE) of 13.7 W with a standard deviation of 22.2 W, while the extreme gradient boosting algorithm (XGBoost) model reduced these values to 9.6 W and 13.0 W, respectively. This corresponds to a reduction in error variance of approximately 40% and an increase in prediction accuracy, with 88.7% of the machine learning (ML) predictions falling within ±20 W of the measured losses, compared to 78.4% for the physical model. Overall, the results demonstrate that the machine learning approach delivers superior predictive precision, while the physics-based formulation remains valuable for interpretability and reliable extrapolation beyond the training domain.
Title and abstract reproduced with permission from CIGRE. Source: CIGRE Session 2026, Paper A2-12607.
Access the full paper
The full paper is available through the official eCIGRE publication platform.