Scholarlink

The thermal characteristics of gas-insulated busbars (GIB) are one of the key indicators determining the current-carrying capacity and reliability of the equipment. This study introduces a physics-informed neural network (PINN) method to reconstruct the temperature fields and inverse heat sources in single-phase GIB units. Finite element (FE) simulations are used to generate sampling points within the two-dimensional GIB solution domain, which are divided into 80% training and 20% testing datasets. A neural network (NN) serves as a surrogate model, regressing velocity, pressure, and temperature based on spatial coordinates. The Navier-Stokes (N-S) equations are incorporated into the loss function using automatic differentiation. A multi-loss function balancing strategy adjusts the weights of different loss terms. The method’s effectiveness is validated against computational fluid dynamics (CFD) simulations, showing a maximum temperature error of 1.09% and a heat flux inversion error of 0.35%. This approach effectively integrates field data with physical laws, making it a promising tool for evaluating the thermal performance and diagnosing overheating in GIB units.

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

Reconstruction of Temperature Field and Inversion of Heat Sources in Gas-Insulated Busbars Based on Physics-Informed Neural Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chen Xiaokun,
  • Xu Xinling,
  • Lai Zekai,
  • Guan Xiangyu

摘要

The thermal characteristics of gas-insulated busbars (GIB) are one of the key indicators determining the current-carrying capacity and reliability of the equipment. This study introduces a physics-informed neural network (PINN) method to reconstruct the temperature fields and inverse heat sources in single-phase GIB units. Finite element (FE) simulations are used to generate sampling points within the two-dimensional GIB solution domain, which are divided into 80% training and 20% testing datasets. A neural network (NN) serves as a surrogate model, regressing velocity, pressure, and temperature based on spatial coordinates. The Navier-Stokes (N-S) equations are incorporated into the loss function using automatic differentiation. A multi-loss function balancing strategy adjusts the weights of different loss terms. The method’s effectiveness is validated against computational fluid dynamics (CFD) simulations, showing a maximum temperature error of 1.09% and a heat flux inversion error of 0.35%. This approach effectively integrates field data with physical laws, making it a promising tool for evaluating the thermal performance and diagnosing overheating in GIB units.