Scholarlink

<p>This study proposes a dynamic prediction method for dissolved hydrogen (H<sub>2</sub>) concentrations in 220&#xa0;kV inverted-type oil-immersed current transformers through a hybrid electro-thermal-fluid multiphysics coupling model integrated with an EEMD-LSTM deep learning framework. Specifically, a three-dimensional finite element analysis model incorporating coupled electromagnetic-thermal-fluid dynamics is developed to simulate H<sub>2</sub> diffusion characteristics under various fault conditions. Systematic multi-temperature gradient simulations quantitatively demonstrate the nonlinear relationship between ambient temperature variations and hydrogen diffusion rates. To address non-stationary concentration signals, ensemble empirical mode decomposition (EEMD) is employed for adaptive noise suppression, effectively extracting intrinsic mode components from high-frequency interference. Subsequently, a gated recurrent unit-optimized LSTM network with temporal feature extraction capability achieves a root mean square error (RMSE) of 0.94% in concentration prediction. Notably, under superimposed noise conditions, the maximum error increases by only 1.01%, ensuring high accuracy. Experimental results confirm that the proposed method outperforms traditional approaches by providing superior tracking of dissolved H<sub>2</sub> variations, enhanced resistance to disturbances, and improved prediction accuracy.</p>

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

Hydrogen concentration prediction in current transformer oil based on multi-physics coupling and EEMD-LSTM

  • 期刊论文

  • 发表日期：

    2026年2月27日

查看全文
  • Hanbo Zheng,
  • Chuanshang Zhang,
  • Xiaopin Deng,
  • Jinrui Kang,
  • Zesen Li

摘要

This study proposes a dynamic prediction method for dissolved hydrogen (H2) concentrations in 220 kV inverted-type oil-immersed current transformers through a hybrid electro-thermal-fluid multiphysics coupling model integrated with an EEMD-LSTM deep learning framework. Specifically, a three-dimensional finite element analysis model incorporating coupled electromagnetic-thermal-fluid dynamics is developed to simulate H2 diffusion characteristics under various fault conditions. Systematic multi-temperature gradient simulations quantitatively demonstrate the nonlinear relationship between ambient temperature variations and hydrogen diffusion rates. To address non-stationary concentration signals, ensemble empirical mode decomposition (EEMD) is employed for adaptive noise suppression, effectively extracting intrinsic mode components from high-frequency interference. Subsequently, a gated recurrent unit-optimized LSTM network with temporal feature extraction capability achieves a root mean square error (RMSE) of 0.94% in concentration prediction. Notably, under superimposed noise conditions, the maximum error increases by only 1.01%, ensuring high accuracy. Experimental results confirm that the proposed method outperforms traditional approaches by providing superior tracking of dissolved H2 variations, enhanced resistance to disturbances, and improved prediction accuracy.