Fast prediction of temperature distributions in oil natural air natural transformers using proper orthogonal decomposition reduced-order data-driven modelling

Abstract

In response to the time-consuming computational fluid dynamics simulations faced in naturally convective oil-immersed transformers, which result from complex models and a high degree of freedom, an innovative reduced-order digital twin prediction model for transformer temperature fields is proposed. This model facilitates fast predictions of transient temperature distributions. Initially, a comprehensive full-order finite element model of transformer temperature distributions is established. Subsequently, a hybrid approach, combining proper orthogonal decomposition (POD)-Galerkin and data-driven techniques, is proposed to create a reduced-order model (ROM). In this model, a Fourier number is utilised as a criterion to select POD training snapshot sets. Subsequently, the dynamic predictive capability of the proposed model under changing operational conditions is validated. Finally, the ROM is employed for fast predictions of temperature field, and its computational errors and time efficiency are compared across diverse operating conditions with full-order models. The research findings confirm the precision, timeliness, and dynamic nature of the reduced-order prediction model, offering a substantial improvement in prediction efficiency and capabilities, all while preserving the accuracy of the digital twin model.