Scholarlink

Power transformers are crucial for stable and efficient electricity transmission within power systems. Managing electromagnetic stress and mechanical vibrations in transformer windings, especially under repeated short-circuit conditions, is vital to prevent failures. This study investigates the use of Physics-Informed Neural Networks (PINNs) to simulate and analyze electromechanical stress. Unlike traditional Finite Element Methods (FEM), which may introduce numerical errors during discretization, PINNs embed physical laws directly into the training process, thereby enhancing accuracy and providing physical explanations in stress predictions. The results indicate that PINNs can accurately predict stress distribution in the central regions of the windings, although they show larger errors at the boundaries due to fewer data points. By improving the distribution of training data and refining boundary conditions, the accuracy of PINNs can be further enhanced. This study highlights the potential of integrating deep learning methods into traditional numerical computations in the power equipment industry, contributing to the digitalization and intelligence of power grid equipment.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Learning Method Based on Physics Informed Neural Networks for the Electromagnetic Stress Simulation in Transformer Windings

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Yuhang Li,
  • Yadong Liu,
  • Yingjie Yan,
  • Jun Wang,
  • Taha Mattar

摘要

Power transformers are crucial for stable and efficient electricity transmission within power systems. Managing electromagnetic stress and mechanical vibrations in transformer windings, especially under repeated short-circuit conditions, is vital to prevent failures. This study investigates the use of Physics-Informed Neural Networks (PINNs) to simulate and analyze electromechanical stress. Unlike traditional Finite Element Methods (FEM), which may introduce numerical errors during discretization, PINNs embed physical laws directly into the training process, thereby enhancing accuracy and providing physical explanations in stress predictions. The results indicate that PINNs can accurately predict stress distribution in the central regions of the windings, although they show larger errors at the boundaries due to fewer data points. By improving the distribution of training data and refining boundary conditions, the accuracy of PINNs can be further enhanced. This study highlights the potential of integrating deep learning methods into traditional numerical computations in the power equipment industry, contributing to the digitalization and intelligence of power grid equipment.